readme: fix broken Swollama link in community integrations (#13370)

Follow up from #12992

Free all streams, and keep the alloc logic aligned across streams.

.gitattributes

@@ -19,6 +19,8 @@ ml/backend/**/*.comp linguist-vendored
 ml/backend/**/*.glsl linguist-vendored
 ml/backend/**/CMakeLists.txt linguist-vendored
 
+app/webview linguist-vendored
+
 llama/build-info.cpp linguist-generated
 ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated
 

.golangci.yaml

@@ -1,77 +1,51 @@
 version: "2"
 linters:
-  default: none
   enable:
     - asasalint
     - bidichk
     - bodyclose
     - containedctx
-    - copyloopvar
-    - errcheck
-    - errorlint
-    - exptostd
     - gocheckcompilerdirectives
-    - gocritic
-    - govet
-    - ineffassign
     - intrange
     - makezero
     - misspell
-    - modernize
     - nilerr
-    - nilnil
     - nolintlint
     - nosprintfhostport
-    - perfsprint
-    - prealloc
-    - sloglint
-    - staticcheck
     - unconvert
-    - unused
-    - usestdlibvars
     - usetesting
     - wastedassign
     - whitespace
+  disable:
+    - errcheck
+    - usestdlibvars
   settings:
-    errcheck:
-      exclude-functions:
-        - fmt.Fprintf
-    perfsprint:
-      strconcat: false
-      concat-loop: false
+    govet:
+      disable:
+        - unusedresult
     staticcheck:
       checks:
         - all
-        # Using a deprecated function, variable, constant or field.
-        # https://staticcheck.dev/docs/checks/#SA1019
+        - -QF* # disable quick fix suggestions
         - -SA1019
-        # Incorrect or missing package comment.
-        # https://staticcheck.dev/docs/checks/#ST1000
-        - -ST1000
-        # Poorly chosen identifier.
-        # https://staticcheck.dev/docs/checks/#ST1003
-        - -ST1003
-        # The documentation of an exported function should start with the function's name.
-        # https://staticcheck.dev/docs/checks/#ST1020
-        - -ST1020
-        # The documentation of an exported type should start with type's name.
-        # https://staticcheck.dev/docs/checks/#ST1021
-        - -ST1021
-        # The documentation of an exported variable or constant should start with variable's name.
-        # https://staticcheck.dev/docs/checks/#ST1022
-        - -ST1022
-    usestdlibvars:
-      http-method: false
-      http-status-code: false
-
+        - -ST1000 # package comment format
+        - -ST1003 # underscores in package names
+        - -ST1005 # error strings should not be capitalized
+        - -ST1012 # error var naming (ErrFoo)
+        - -ST1016 # receiver name consistency
+        - -ST1020 # comment on exported function format
+        - -ST1021 # comment on exported type format
+        - -ST1022 # comment on exported var format
+        - -ST1023 # omit type from declaration
+severity:
+  default: error
+  rules:
+    - linters:
+        - gofmt
+        - goimports
+        - intrange
+      severity: info
 formatters:
   enable:
-    - gci
     - gofmt
     - gofumpt
-  settings:
-    gci:
-      sections:
-        - standard
-        - default
-        - localmodule

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=3cfa9c3f125763305b4226bc032f1954f08990dc
+FETCH_HEAD=7f8ef50cce40e3e7e4526a3696cb45658190e69a
 
 .PHONY: help
 help:

README.md

@@ -555,7 +555,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Parakeet](https://github.com/parakeet-nest/parakeet) is a GoLang library, made to simplify the development of small generative AI applications with Ollama.
 - [Haverscript](https://github.com/andygill/haverscript) with [examples](https://github.com/andygill/haverscript/tree/main/examples)
 - [Ollama for Swift](https://github.com/mattt/ollama-swift)
-- [Swollama for Swift](https://github.com/marcusziade/Swollama) with [DocC](https://marcusziade.github.io/Swollama/documentation/swollama/)
+- [Swollama for Swift]([https://github.com/marcusziade/Swollama](https://github.com/guitaripod/Swollama) with [DocC]( https://guitaripod.github.io/Swollama/documentation/swollama)
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in a unified API)

api/client.go

@@ -226,7 +226,14 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 
 		bts := scanner.Bytes()
 		if err := json.Unmarshal(bts, &errorResponse); err != nil {
-			return fmt.Errorf("unmarshal: %w", err)
+			if response.StatusCode >= http.StatusBadRequest {
+				return StatusError{
+					StatusCode:   response.StatusCode,
+					Status:       response.Status,
+					ErrorMessage: string(bts),
+				}
+			}
+			return errors.New(string(bts))
 		}
 
 		if response.StatusCode == http.StatusUnauthorized {

api/client_test.go

@@ -55,6 +55,7 @@ func TestClientFromEnvironment(t *testing.T) {
 type testError struct {
 	message    string
 	statusCode int
+	raw        bool // if true, write message as-is instead of JSON encoding
 }
 
 func (e testError) Error() string {
@@ -111,6 +112,20 @@ func TestClientStream(t *testing.T) {
 				},
 			},
 		},
+		{
+			name: "plain text error response",
+			responses: []any{
+				"internal server error",
+			},
+			wantErr: "internal server error",
+		},
+		{
+			name: "HTML error page",
+			responses: []any{
+				"<html><body>404 Not Found</body></html>",
+			},
+			wantErr: "404 Not Found",
+		},
 	}
 
 	for _, tc := range testCases {
@@ -135,6 +150,12 @@ func TestClientStream(t *testing.T) {
 						return
 					}
 
+					if str, ok := resp.(string); ok {
+						fmt.Fprintln(w, str)
+						flusher.Flush()
+						continue
+					}
+
 					if err := json.NewEncoder(w).Encode(resp); err != nil {
 						t.Fatalf("failed to encode response: %v", err)
 					}
@@ -173,9 +194,10 @@ func TestClientStream(t *testing.T) {
 
 func TestClientDo(t *testing.T) {
 	testCases := []struct {
-		name     string
-		response any
-		wantErr  string
+		name           string
+		response       any
+		wantErr        string
+		wantStatusCode int
 	}{
 		{
 			name: "immediate error response",
@@ -183,7 +205,8 @@ func TestClientDo(t *testing.T) {
 				message:    "test error message",
 				statusCode: http.StatusBadRequest,
 			},
-			wantErr: "test error message",
+			wantErr:        "test error message",
+			wantStatusCode: http.StatusBadRequest,
 		},
 		{
 			name: "server error response",
@@ -191,7 +214,8 @@ func TestClientDo(t *testing.T) {
 				message:    "internal error",
 				statusCode: http.StatusInternalServerError,
 			},
-			wantErr: "internal error",
+			wantErr:        "internal error",
+			wantStatusCode: http.StatusInternalServerError,
 		},
 		{
 			name: "successful response",
@@ -203,6 +227,26 @@ func TestClientDo(t *testing.T) {
 				Success: true,
 			},
 		},
+		{
+			name: "plain text error response",
+			response: testError{
+				message:    "internal server error",
+				statusCode: http.StatusInternalServerError,
+				raw:        true,
+			},
+			wantErr:        "internal server error",
+			wantStatusCode: http.StatusInternalServerError,
+		},
+		{
+			name: "HTML error page",
+			response: testError{
+				message:    "<html><body>404 Not Found</body></html>",
+				statusCode: http.StatusNotFound,
+				raw:        true,
+			},
+			wantErr:        "<html><body>404 Not Found</body></html>",
+			wantStatusCode: http.StatusNotFound,
+		},
 	}
 
 	for _, tc := range testCases {
@@ -210,11 +254,16 @@ func TestClientDo(t *testing.T) {
 			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 				if errResp, ok := tc.response.(testError); ok {
 					w.WriteHeader(errResp.statusCode)
-					err := json.NewEncoder(w).Encode(map[string]string{
-						"error": errResp.message,
-					})
-					if err != nil {
-						t.Fatal("failed to encode error response:", err)
+					if !errResp.raw {
+						err := json.NewEncoder(w).Encode(map[string]string{
+							"error": errResp.message,
+						})
+						if err != nil {
+							t.Fatal("failed to encode error response:", err)
+						}
+					} else {
+						// Write raw message (simulates non-JSON error responses)
+						fmt.Fprint(w, errResp.message)
 					}
 					return
 				}
@@ -241,6 +290,15 @@ func TestClientDo(t *testing.T) {
 				if err.Error() != tc.wantErr {
 					t.Errorf("error message mismatch: got %q, want %q", err.Error(), tc.wantErr)
 				}
+				if tc.wantStatusCode != 0 {
+					if statusErr, ok := err.(StatusError); ok {
+						if statusErr.StatusCode != tc.wantStatusCode {
+							t.Errorf("status code mismatch: got %d, want %d", statusErr.StatusCode, tc.wantStatusCode)
+						}
+					} else {
+						t.Errorf("expected StatusError, got %T", err)
+					}
+				}
 				return
 			}
 

api/examples/chat/main.go

@@ -15,19 +15,19 @@ func main() {
 	}
 
 	messages := []api.Message{
-		api.Message{
+		{
 			Role:    "system",
 			Content: "Provide very brief, concise responses",
 		},
-		api.Message{
+		{
 			Role:    "user",
 			Content: "Name some unusual animals",
 		},
-		api.Message{
+		{
 			Role:    "assistant",
 			Content: "Monotreme, platypus, echidna",
 		},
-		api.Message{
+		{
 			Role:    "user",
 			Content: "which of these is the most dangerous?",
 		},

app/ui/app/src/api.ts

@@ -204,12 +204,10 @@ export async function* sendMessage(
     data: uint8ArrayToBase64(att.data),
   }));
 
-  // Only send think parameter when actually requesting thinking
-  // Don't send false as it causes issues with some providers
+  // Send think parameter when it's explicitly set (true, false, or a non-empty string).
   const shouldSendThink =
     think !== undefined &&
-    ((typeof think === "boolean" && think) ||
-      (typeof think === "string" && think !== ""));
+    (typeof think === "boolean" || (typeof think === "string" && think !== ""));
 
   const response = await fetch(`${API_BASE}/api/v1/chat/${chatId}`, {
     method: "POST",

cmd/bench/bench.go

@@ -48,8 +48,8 @@ func OutputMetrics(w io.Writer, format string, metrics []Metrics, verbose bool)
 	case "benchstat":
 		if verbose {
 			printHeader := func() {
-				fmt.Printf("sysname: %s\n", runtime.GOOS)
-				fmt.Printf("machine: %s\n", runtime.GOARCH)
+				fmt.Fprintf(w, "sysname: %s\n", runtime.GOOS)
+				fmt.Fprintf(w, "machine: %s\n", runtime.GOARCH)
 			}
 			once.Do(printHeader)
 		}
@@ -147,6 +147,17 @@ func BenchmarkChat(fOpt flagOptions) error {
 		return err
 	}
 
+	var out io.Writer = os.Stdout
+	if fOpt.outputFile != nil && *fOpt.outputFile != "" {
+		f, err := os.OpenFile(*fOpt.outputFile, os.O_CREATE|os.O_WRONLY, 0o644)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "ERROR: cannot open output file %s: %v\n", *fOpt.outputFile, err)
+			return err
+		}
+		defer f.Close()
+		out = f
+	}
+
 	for _, model := range models {
 		for range *fOpt.epochs {
 			options := make(map[string]interface{})
@@ -241,13 +252,14 @@ func BenchmarkChat(fOpt flagOptions) error {
 				},
 			}
 
-			OutputMetrics(os.Stdout, *fOpt.format, metrics, *fOpt.verbose)
+			OutputMetrics(out, *fOpt.format, metrics, *fOpt.verbose)
 
 			if *fOpt.keepAlive > 0 {
 				time.Sleep(time.Duration(*fOpt.keepAlive*float64(time.Second)) + 200*time.Millisecond)
 			}
 		}
 	}
+
 	return nil
 }
 

cmd/cmd.go

@@ -1430,7 +1430,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		latest.Summary()
 	}
 
-	return &api.Message{Role: role, Content: fullResponse.String()}, nil
+	return &api.Message{Role: role, Thinking: thinkingContent.String(), Content: fullResponse.String()}, nil
 }
 
 func generate(cmd *cobra.Command, opts runOptions) error {

cmd/eval/README.md

@@ -1,50 +0,0 @@
-# eval
-
-Evaluation tool for testing Ollama models.
-
-## Usage
-
-Run all tests:
-
-```bash
-go run . -model llama3.2:latest
-```
-
-Run specific suite:
-
-```bash
-go run . -model llama3.2:latest -suite tool-calling-basic -v
-```
-
-List available suites:
-
-```bash
-go run . -list
-```
-
-## Adding Tests
-
-Edit `suites.go` to add new test suites. Each test needs:
-
-- `Name`: test identifier
-- `Prompt`: what to send to the model
-- `Check`: function to validate the response
-
-Example:
-
-```go
-{
-    Name:   "my-test",
-    Prompt: "What is 2+2?",
-    Check:  Contains("4"),
-}
-```
-
-Available check functions:
-
-- `HasResponse()` - response is non-empty
-- `Contains(s)` - response contains substring
-- `CallsTool(name)` - model called specific tool
-- `NoTools()` - model called no tools
-- `MinTools(n)` - model called at least n tools
-- `All(checks...)` - all checks pass

cmd/eval/eval.go

@@ -1,151 +0,0 @@
-package main
-
-import (
-	"context"
-	"strings"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// Test is a single evaluation test
-type Test struct {
-	Name    string
-	Prompt  string
-	System  string
-	Tools   []api.Tool
-	Think   bool
-	Options map[string]any
-	Check   func(response string, tools []api.ToolCall) bool
-}
-
-// Suite is a collection of tests
-type Suite struct {
-	Name  string
-	Tests []Test
-}
-
-// Result holds test execution results
-type Result struct {
-	Name      string
-	Passed    bool
-	Error     error
-	Duration  time.Duration
-	Response  string
-	Tools     []string
-	ToolCalls []api.ToolCall
-	Thinking  bool
-}
-
-// Run executes a test against a model
-func Run(ctx context.Context, client *api.Client, model string, test Test) Result {
-	result := Result{Name: test.Name}
-
-	req := &api.ChatRequest{
-		Model: model,
-		Messages: []api.Message{
-			{Role: "user", Content: test.Prompt},
-		},
-		Options: test.Options,
-	}
-
-	if test.System != "" {
-		req.Messages = append([]api.Message{
-			{Role: "system", Content: test.System},
-		}, req.Messages...)
-	}
-
-	if len(test.Tools) > 0 {
-		req.Tools = test.Tools
-	}
-
-	if test.Think {
-		req.Think = &api.ThinkValue{Value: true}
-	}
-
-	var resp strings.Builder
-	var toolCalls []api.ToolCall
-
-	start := time.Now()
-	err := client.Chat(ctx, req, func(r api.ChatResponse) error {
-		resp.WriteString(r.Message.Content)
-		if r.Message.Thinking != "" {
-			result.Thinking = true
-		}
-		toolCalls = append(toolCalls, r.Message.ToolCalls...)
-		return nil
-	})
-	result.Duration = time.Since(start)
-
-	if err != nil {
-		result.Error = err
-		return result
-	}
-
-	result.Response = resp.String()
-	result.Tools = uniqueToolNames(toolCalls)
-	result.ToolCalls = toolCalls
-	result.Passed = test.Check(result.Response, toolCalls)
-
-	return result
-}
-
-func uniqueToolNames(calls []api.ToolCall) []string {
-	seen := make(map[string]bool)
-	var names []string
-	for _, c := range calls {
-		if !seen[c.Function.Name] {
-			seen[c.Function.Name] = true
-			names = append(names, c.Function.Name)
-		}
-	}
-	return names
-}
-
-// Check functions for common test patterns
-
-func HasResponse() func(string, []api.ToolCall) bool {
-	return func(resp string, _ []api.ToolCall) bool {
-		return strings.TrimSpace(resp) != ""
-	}
-}
-
-func Contains(s string) func(string, []api.ToolCall) bool {
-	return func(resp string, _ []api.ToolCall) bool {
-		return strings.Contains(strings.ToLower(resp), strings.ToLower(s))
-	}
-}
-
-func CallsTool(name string) func(string, []api.ToolCall) bool {
-	return func(_ string, tools []api.ToolCall) bool {
-		for _, t := range tools {
-			if t.Function.Name == name {
-				return true
-			}
-		}
-		return false
-	}
-}
-
-func NoTools() func(string, []api.ToolCall) bool {
-	return func(_ string, tools []api.ToolCall) bool {
-		return len(tools) == 0
-	}
-}
-
-func MinTools(n int) func(string, []api.ToolCall) bool {
-	return func(_ string, tools []api.ToolCall) bool {
-		return len(tools) >= n
-	}
-}
-
-func All(checks ...func(string, []api.ToolCall) bool) func(string, []api.ToolCall) bool {
-	return func(resp string, tools []api.ToolCall) bool {
-		for _, check := range checks {
-			if !check(resp, tools) {
-				return false
-			}
-		}
-		return true
-	}
-}

cmd/eval/main.go

@@ -1,217 +0,0 @@
-package main
-
-import (
-	"context"
-	"encoding/json"
-	"flag"
-	"fmt"
-	"os"
-	"strings"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-func main() {
-	model := flag.String("model", "", "model to evaluate")
-	suite := flag.String("suite", "", "comma-separated list of suites to run (empty runs all)")
-	list := flag.Bool("list", false, "list available suites")
-	verbose := flag.Bool("v", false, "verbose output")
-	timeout := flag.Int("timeout", 60, "timeout per test in seconds")
-	export := flag.String("export", "eval-results.json", "export results to file")
-	flag.Parse()
-
-	if *list {
-		for _, s := range suites {
-			fmt.Printf("%s (%d tests)\n", s.Name, len(s.Tests))
-		}
-		return
-	}
-
-	if *model == "" {
-		fmt.Fprintf(os.Stderr, "error: -model parameter is required\n")
-		os.Exit(1)
-	}
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "error: %v\n", err)
-		os.Exit(1)
-	}
-
-	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
-	if err := client.Heartbeat(ctx); err != nil {
-		cancel()
-		fmt.Fprintf(os.Stderr, "error: cannot connect to ollama\n")
-		os.Exit(1)
-	}
-	cancel()
-
-	selected := suites
-	if *suite != "" {
-		suiteNames := strings.Split(*suite, ",")
-		selected = []Suite{}
-		var notFound []string
-
-		for _, name := range suiteNames {
-			name = strings.TrimSpace(name)
-			if name == "" {
-				continue
-			}
-
-			found := false
-			for _, s := range suites {
-				if s.Name == name {
-					selected = append(selected, s)
-					found = true
-					break
-				}
-			}
-			if !found {
-				notFound = append(notFound, name)
-			}
-		}
-
-		if len(notFound) > 0 {
-			fmt.Fprintf(os.Stderr, "error: suite(s) not found: %s\n", strings.Join(notFound, ", "))
-			os.Exit(1)
-		}
-	}
-
-	var results []Result
-	for _, s := range selected {
-		if *verbose {
-			fmt.Printf("\n%s (%d tests)\n", s.Name, len(s.Tests))
-		}
-		for i, test := range s.Tests {
-			if test.Options == nil {
-				test.Options = map[string]any{"temperature": 0.1}
-			}
-			if test.Check == nil {
-				test.Check = HasResponse()
-			}
-
-			if *verbose {
-				fmt.Printf("  [%d/%d] %s... ", i+1, len(s.Tests), test.Name)
-			}
-
-			ctx, cancel := context.WithTimeout(context.Background(), time.Duration(*timeout)*time.Second)
-			result := Run(ctx, client, *model, test)
-			cancel()
-
-			results = append(results, result)
-
-			if *verbose {
-				if result.Error != nil {
-					fmt.Printf("ERROR: %v\n", result.Error)
-				} else if result.Passed {
-					fmt.Printf("PASS (%.2fs)", result.Duration.Seconds())
-					if len(result.Tools) > 0 || result.Thinking {
-						fmt.Printf(" [")
-						if len(result.Tools) > 0 {
-							fmt.Printf("tools: %s", strings.Join(result.Tools, ","))
-						}
-						if result.Thinking {
-							if len(result.Tools) > 0 {
-								fmt.Printf(", ")
-							}
-							fmt.Printf("thinking")
-						}
-						fmt.Printf("]")
-					}
-					fmt.Println()
-
-					// Print tool calls with details
-					if len(result.ToolCalls) > 0 {
-						fmt.Printf("    Tool Calls:\n")
-						for _, tc := range result.ToolCalls {
-							argsJSON, _ := json.Marshal(tc.Function.Arguments)
-							fmt.Printf("      - %s: %s\n", tc.Function.Name, string(argsJSON))
-						}
-					}
-
-					// Print response if there is one
-					if result.Response != "" {
-						fmt.Printf("    Response: %s\n", result.Response)
-					}
-				} else {
-					fmt.Printf("FAIL (%.2fs)\n", result.Duration.Seconds())
-
-					// Print tool calls with details even on failure
-					if len(result.ToolCalls) > 0 {
-						fmt.Printf("    Tool Calls:\n")
-						for _, tc := range result.ToolCalls {
-							argsJSON, _ := json.Marshal(tc.Function.Arguments)
-							fmt.Printf("      - %s: %s\n", tc.Function.Name, string(argsJSON))
-						}
-					}
-
-					// Print response even on failure
-					if result.Response != "" {
-						fmt.Printf("    Response: %s\n", result.Response)
-					}
-				}
-			}
-		}
-	}
-
-	printSummary(results)
-
-	if *export != "" {
-		if err := writeJSON(*export, results); err != nil {
-			fmt.Fprintf(os.Stderr, "warning: export failed: %v\n", err)
-		} else if *verbose {
-			fmt.Printf("\nResults: %s\n", *export)
-		}
-	}
-
-	if anyFailed(results) {
-		os.Exit(1)
-	}
-}
-
-func printSummary(results []Result) {
-	var passed, failed, errors int
-	for _, r := range results {
-		if r.Error != nil {
-			errors++
-		} else if r.Passed {
-			passed++
-		} else {
-			failed++
-		}
-	}
-
-	total := len(results)
-	rate := 0.0
-	if total > 0 {
-		rate = float64(passed) / float64(total) * 100
-	}
-
-	fmt.Printf("\n%d/%d passed (%.1f%%)", passed, total, rate)
-	if errors > 0 {
-		fmt.Printf(", %d errors", errors)
-	}
-	fmt.Println()
-}
-
-func anyFailed(results []Result) bool {
-	for _, r := range results {
-		if !r.Passed || r.Error != nil {
-			return true
-		}
-	}
-	return false
-}
-
-func writeJSON(path string, results []Result) error {
-	f, err := os.Create(path)
-	if err != nil {
-		return err
-	}
-	defer f.Close()
-
-	enc := json.NewEncoder(f)
-	enc.SetIndent("", "  ")
-	return enc.Encode(results)
-}

cmd/eval/suites.go

@@ -1,178 +0,0 @@
-package main
-
-import "github.com/ollama/ollama/api"
-
-var suites = []Suite{
-	{
-		Name: "basic-qa",
-		Tests: []Test{
-			{
-				Name:   "simple-math",
-				Prompt: "What is 2+2? Reply with just the number.",
-				Check:  Contains("4"),
-			},
-			{
-				Name:   "capital-city",
-				Prompt: "What is the capital of France? Reply with just the city name.",
-				Check:  Contains("Paris"),
-			},
-			{
-				Name:   "greeting",
-				Prompt: "Say hello",
-				Check:  HasResponse(),
-			},
-		},
-	},
-	{
-		Name: "reasoning",
-		Tests: []Test{
-			{
-				Name:   "logic-puzzle",
-				Prompt: "If all roses are flowers and some flowers fade quickly, can we conclude that some roses fade quickly? Answer yes or no.",
-				Check:  Contains("no"),
-			},
-			{
-				Name:   "counting",
-				Prompt: "How many letters are in the word 'HELLO'?",
-				Check:  Contains("5"),
-			},
-		},
-	},
-	{
-		Name: "instruction-following",
-		Tests: []Test{
-			{
-				Name:   "json-output",
-				Prompt: "Reply with a JSON object containing a 'status' field set to 'ok'.",
-				Check:  All(Contains("status"), Contains("ok")),
-			},
-			{
-				Name:   "system-prompt",
-				Prompt: "What is your name?",
-				System: "You are a helpful assistant named TestBot. When asked your name, always respond with 'TestBot'.",
-				Check:  Contains("TestBot"),
-			},
-		},
-	},
-	{
-		Name: "tool-calling-basic",
-		Tests: []Test{
-			{
-				Name:   "single-tool",
-				Prompt: "What's the weather like in San Francisco?",
-				Tools:  []api.Tool{weatherTool},
-				Check:  CallsTool("get_weather"),
-			},
-			{
-				Name:   "tool-selection",
-				Prompt: "What time is it in Tokyo?",
-				Tools:  []api.Tool{weatherTool, timeTool},
-				Check:  CallsTool("get_time"),
-			},
-			{
-				Name:   "no-tool-needed",
-				Prompt: "What is 2+2?",
-				Tools:  []api.Tool{weatherTool, timeTool},
-				Check:  NoTools(),
-			},
-		},
-	},
-	{
-		Name: "tool-calling-advanced",
-		Tests: []Test{
-			{
-				Name:   "parallel-calls",
-				Prompt: "Get the weather in both New York and Los Angeles.",
-				Tools:  []api.Tool{weatherTool},
-				Check:  All(CallsTool("get_weather"), MinTools(2)),
-			},
-			{
-				Name:   "multi-param",
-				Prompt: "Search for Italian restaurants with prices between $20 and $40.",
-				Tools:  []api.Tool{restaurantTool},
-				Check:  CallsTool("search_restaurants"),
-			},
-		},
-	},
-	{
-		Name: "tool-calling-thinking",
-		Tests: []Test{
-			{
-				Name:   "thinking-before-tool",
-				Prompt: "I need to know the weather in Paris before I decide what to pack.",
-				Tools:  []api.Tool{weatherTool},
-				Think:  true,
-				Check:  CallsTool("get_weather"),
-			},
-			{
-				Name:   "thinking-multi-tool",
-				Prompt: "I'm planning a trip to London. I need to know what time it is there and what the weather is like.",
-				Tools:  []api.Tool{weatherTool, timeTool},
-				Think:  true,
-				Check:  MinTools(1),
-			},
-		},
-	},
-}
-
-var weatherTool = api.Tool{
-	Type: "function",
-	Function: api.ToolFunction{
-		Name:        "get_weather",
-		Description: "Get the current weather in a given location",
-		Parameters: api.ToolFunctionParameters{
-			Type:     "object",
-			Required: []string{"location"},
-			Properties: map[string]api.ToolProperty{
-				"location": {
-					Type:        api.PropertyType{"string"},
-					Description: "The city and state",
-				},
-			},
-		},
-	},
-}
-
-var timeTool = api.Tool{
-	Type: "function",
-	Function: api.ToolFunction{
-		Name:        "get_time",
-		Description: "Get the current time in a timezone",
-		Parameters: api.ToolFunctionParameters{
-			Type:     "object",
-			Required: []string{"timezone"},
-			Properties: map[string]api.ToolProperty{
-				"timezone": {
-					Type:        api.PropertyType{"string"},
-					Description: "The timezone name",
-				},
-			},
-		},
-	},
-}
-
-var restaurantTool = api.Tool{
-	Type: "function",
-	Function: api.ToolFunction{
-		Name:        "search_restaurants",
-		Description: "Search for restaurants",
-		Parameters: api.ToolFunctionParameters{
-			Type:     "object",
-			Required: []string{"cuisine"},
-			Properties: map[string]api.ToolProperty{
-				"cuisine": {
-					Type:        api.PropertyType{"string"},
-					Description: "Type of cuisine",
-				},
-				"min_price": {
-					Type:        api.PropertyType{"number"},
-					Description: "Minimum price",
-				},
-				"max_price": {
-					Type:        api.PropertyType{"number"},
-					Description: "Maximum price",
-				},
-			},
-		},
-	},
-}

convert/convert.go

@@ -208,6 +208,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &gptossModel{}
 	case "DeepseekOCRForCausalLM":
 		conv = &deepseekocr{}
+	case "DeepseekV3ForCausalLM":
+		conv = &deepseek2Model{}
 	default:
 		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_deepseek2.go

@@ -0,0 +1,173 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type deepseek2Model struct {
+	ModelParameters               // architectures, vocab_size
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	ScoringFunc            string `json:"scoring_func"`
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+
+	RopeScaling struct {
+		Factor                        float32 `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		Type                          string  `json:"type"`
+		MScaleAllDim                  float32 `json:"mscale_all_dim"`
+	} `json:"rope_scaling"`
+
+	Architecture string
+}
+
+func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "deepseek2"
+	kv["general.type"] = "model"
+	kv["deepseek2.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["deepseek2.attention.head_count"] = numHeads
+	kv["deepseek2.attention.head_count_kv"] = numKVHeads
+	kv["deepseek2.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["deepseek2.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["deepseek2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["deepseek2.attention.q_lora_rank"] = p.QLoraRank
+	kv["deepseek2.attention.value_length"] = p.VHeadDim
+	kv["deepseek2.context_length"] = p.MaxPositionEmbeddings
+	kv["deepseek2.embedding_length"] = p.HiddenSize
+	kv["deepseek2.expert_count"] = p.ExpertCount
+	kv["deepseek2.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["deepseek2.expert_shared_count"] = p.ExpertSharedCount
+
+	var scoringFunc uint32
+	switch p.ScoringFunc {
+	case "softmax":
+		// not currently supported in the model, but needed for Deepseek-OCR
+		scoringFunc = 1
+	case "sigmoid":
+		scoringFunc = 2
+	}
+	kv["deepseek2.expert_gating_func"] = scoringFunc
+	kv["deepseek2.expert_used_count"] = p.ExpertUsedCount
+	kv["deepseek2.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["deepseek2.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["deepseek2.feed_forward_length"] = p.IntermediateSize
+	kv["deepseek2.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["deepseek2.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["deepseek2.rope.freq_base"] = cmp.Or(p.RopeTheta, 10000.0)
+	kv["deepseek2.rope.scaling.factor"] = p.RopeScaling.Factor
+	kv["deepseek2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
+	kv["deepseek2.rope.scaling.type"] = p.RopeScaling.Type
+	kv["deepseek2.rope.scaling.yarn_log_multiplier"] = 0.1 * p.RopeScaling.MScaleAllDim
+
+	kv["tokenizer.ggml.pre"] = "deepseek-v3"
+
+	return kv
+}
+
+func (p *deepseek2Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"language_model.", "",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *deepseek2Model) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_mistral.go

@@ -29,6 +29,15 @@ type mistral3Model struct {
 		SlidingWindow         *uint32 `json:"sliding_window"`
 		HiddenAct             string  `json:"hidden_act"`
 		VocabSize             uint32  `json:"vocab_size"`
+		RopeParameters        struct {
+			BetaFast                  float32 `json:"beta_fast"`
+			BetaSlow                  float32 `json:"beta_slow"`
+			Factor                    float32 `json:"factor"`
+			ScalingBeta               float32 `json:"llama_4_scaling_beta"`
+			OrigMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+			RopeType                  string  `json:"rope_type"`
+			RopeTheta                 float32 `json:"rope_theta"`
+		} `json:"rope_parameters"`
 	} `json:"text_config"`
 	VisionModel struct {
 		NumAttentionHeads uint32  `json:"num_attention_heads"`
@@ -61,8 +70,13 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
 	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
 	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
-	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
-	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
+	kv["mistral3.rope.dimension_count"] = cmp.Or(p.TextModel.HeadDim, p.TextModel.HiddenSize/p.TextModel.NumAttentionHeads)
+	kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
+
+	if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
+		kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
+		kv["mistral3.rope.scaling_beta"] = p.TextModel.RopeParameters.ScalingBeta
+	}
 
 	// Vision configuration
 	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers

discover/runner.go

@@ -65,6 +65,7 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 		}
 
 		slog.Info("discovering available GPUs...")
+		detectIncompatibleLibraries()
 
 		// Warn if any user-overrides are set which could lead to incorrect GPU discovery
 		overrideWarnings()
@@ -98,6 +99,9 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 					continue
 				} else if jetpack != "" && filepath.Base(dir) != "cuda_"+jetpack {
 					continue
+				} else if jetpack == "" && strings.Contains(filepath.Base(dir), "cuda_jetpack") {
+					slog.Debug("jetpack not detected (set JETSON_JETPACK or OLLAMA_LLM_LIBRARY to override), skipping", "libDir", dir)
+					continue
 				} else if !envconfig.EnableVulkan() && strings.Contains(filepath.Base(dir), "vulkan") {
 					slog.Info("experimental Vulkan support disabled.  To enable, set OLLAMA_VULKAN=1")
 					continue
@@ -143,7 +147,7 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			wg.Add(1)
 			go func(i int) {
 				defer wg.Done()
-				extraEnvs := ml.GetVisibleDevicesEnv(devices[i : i+1])
+				extraEnvs := ml.GetVisibleDevicesEnv(devices[i:i+1], true)
 				devices[i].AddInitValidation(extraEnvs)
 				if len(bootstrapDevices(ctx2ndPass, devices[i].LibraryPath, extraEnvs)) == 0 {
 					slog.Debug("filtering device which didn't fully initialize",
@@ -329,7 +333,8 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			defer cancel()
 
 			// Apply any dev filters to avoid re-discovering unsupported devices, and get IDs correct
-			devFilter := ml.GetVisibleDevicesEnv(devices)
+			// We avoid CUDA filters here to keep ROCm from failing to discover GPUs in a mixed environment
+			devFilter := ml.GetVisibleDevicesEnv(devices, false)
 
 			for dir := range libDirs {
 				updatedDevices := bootstrapDevices(ctx, []string{ml.LibOllamaPath, dir}, devFilter)
@@ -484,3 +489,16 @@ func overrideWarnings() {
 		slog.Warn("if GPUs are not correctly discovered, unset and try again")
 	}
 }
+
+func detectIncompatibleLibraries() {
+	if runtime.GOOS != "windows" {
+		return
+	}
+	basePath, err := exec.LookPath("ggml-base.dll")
+	if err != nil || basePath == "" {
+		return
+	}
+	if !strings.HasPrefix(basePath, ml.LibOllamaPath) {
+		slog.Warn("potentially incompatible library detected in PATH", "location", basePath)
+	}
+}

docs/capabilities/tool-calling.mdx

@@ -15,7 +15,7 @@ Also known as "single-shot" tool calling.
     ```shell
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
-      "messages": [{"role": "user", "content": "What's the temperature in New York?"}],
+      "messages": [{"role": "user", "content": "What is the temperature in New York?"}],
       "stream": false,
       "tools": [
         {
@@ -41,7 +41,7 @@ Also known as "single-shot" tool calling.
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
       "messages": [
-        {"role": "user", "content": "What's the temperature in New York?"},
+        {"role": "user", "content": "What is the temperature in New York?"},
         {
           "role": "assistant",
           "tool_calls": [
@@ -90,7 +90,7 @@ Also known as "single-shot" tool calling.
       }
       return temperatures.get(city, "Unknown")
 
-    messages = [{"role": "user", "content": "What's the temperature in New York?"}]
+    messages = [{"role": "user", "content": "What is the temperature in New York?"}]
 
     # pass functions directly as tools in the tools list or as a JSON schema
     response = chat(model="qwen3", messages=messages, tools=[get_temperature], think=True)
@@ -146,7 +146,7 @@ Also known as "single-shot" tool calling.
       },
     ]
 
-    const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
+    const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
 
     const response = await ollama.chat({
       model: 'qwen3',
@@ -609,7 +609,7 @@ def get_temperature(city: str) -> str:
   return temperatures.get(city, 'Unknown')
 
 
-messages = [{'role': 'user', 'content': "What's the temperature in New York?"}]
+messages = [{'role': 'user', 'content': "What is the temperature in New York?"}]
 
 while True:
   stream = chat(
@@ -684,7 +684,7 @@ const getTemperatureTool = {
 }
 
 async function agentLoop() {
-  const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
+  const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
 
   while (true) {
     const stream = await ollama.chat({

docs/development.md

@@ -49,6 +49,8 @@ Install prerequisites:
     - [Ninja](https://github.com/ninja-build/ninja/releases)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
+- (Optional) VULKAN GPU support
+    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
 
 Then, configure and build the project:
 
@@ -57,6 +59,17 @@ cmake -B build
 cmake --build build --config Release
 ```
 
+> Building for Vulkan requires VULKAN_SDK environment variable:
+> 
+> PowerShell
+> ```powershell
+> $env:VULKAN_SDK="C:\VulkanSDK\<version>"
+> ```
+> CMD
+> ```cmd
+> set VULKAN_SDK=C:\VulkanSDK\<version>
+> ```
+
 > [!IMPORTANT]
 > Building for ROCm requires additional flags:
 > ```
@@ -65,6 +78,7 @@ cmake --build build --config Release
 > ```
 
 
+
 Lastly, run Ollama:
 
 ```shell
@@ -84,7 +98,9 @@ Install prerequisites:
     - [ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/quick-start.html)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads)
-
+- (Optional) VULKAN GPU support
+    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
+    - Or install via package manager: `sudo apt install vulkan-sdk` (Ubuntu/Debian) or `sudo dnf install vulkan-sdk` (Fedora/CentOS)
 > [!IMPORTANT]
 > Ensure prerequisites are in `PATH` before running CMake.
 

docs/faq.mdx

@@ -57,8 +57,13 @@ ollama ps
 ```
 
 <Info>
-  **Output**: ``` NAME ID SIZE PROCESSOR UNTIL llama3:70b bcfb190ca3a7 42 GB
-  100% GPU 4 minutes from now ```
+
+**Output**:
+
+```
+NAME        ID            SIZE    PROCESSOR   UNTIL
+llama3:70b  bcfb190ca3a7  42 GB   100% GPU    4 minutes from now
+```
 </Info>
 
 The `Processor` column will show which memory the model was loaded in to:
@@ -385,4 +390,4 @@ Ollama for Windows and macOS register as a login item during installation.  You
 - In `Task Manager` go to the `Startup apps` tab, search for `ollama` then click `Disable`
 
 **MacOS**
-- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.

fs/ggml/ggml.go

@@ -831,6 +831,7 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"gemma3",
 		"gptoss", "gpt-oss",
+		"mistral3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

fs/ggml/gguf.go

@@ -597,6 +597,10 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 
 	var err error
 	switch v := v.(type) {
+	case int32:
+		err = writeGGUF(ws, ggufTypeInt32, v)
+	case int64:
+		err = writeGGUF(ws, ggufTypeInt64, v)
 	case uint32, FileType:
 		err = writeGGUF(ws, ggufTypeUint32, v)
 	case uint64:
@@ -611,6 +615,10 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeInt32, v)
 	case *array[int32]:
 		err = writeGGUFArray(ws, ggufTypeInt32, v.values)
+	case []int64:
+		err = writeGGUFArray(ws, ggufTypeInt64, v)
+	case *array[int64]:
+		err = writeGGUFArray(ws, ggufTypeInt64, v.values)
 	case []uint32:
 		err = writeGGUFArray(ws, ggufTypeUint32, v)
 	case *array[uint32]:

fs/ggml/gguf_test.go

@@ -42,6 +42,10 @@ func TestWriteGGUF(t *testing.T) {
 				"general.architecture": "test",
 				"general.alignment":    uint32(16),
 				"test.key":             "value",
+				"test.int32_key":       int32(-42),
+				"test.int64_key":       int64(-9223372036854775808),
+				"test.int32_array":     []int32{-1, 0, 1, 2147483647, -2147483648},
+				"test.int64_array":     []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808},
 				"attention.key":        "value2",
 				"tokenizer.key":        "value3",
 				"adapter.key":          "value4",
@@ -55,7 +59,7 @@ func TestWriteGGUF(t *testing.T) {
 			}
 			defer r.Close()
 
-			ff, err := Decode(r, 0)
+			ff, err := Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -65,15 +69,19 @@ func TestWriteGGUF(t *testing.T) {
 				"general.alignment":       uint32(16),
 				"general.parameter_count": uint64(54),
 				"test.key":                "value",
+				"test.int32_key":          int32(-42),
+				"test.int64_key":          int64(-9223372036854775808),
+				"test.int32_array":        &array[int32]{size: 5, values: []int32{-1, 0, 1, 2147483647, -2147483648}},
+				"test.int64_array":        &array[int64]{size: 5, values: []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808}},
 				"test.attention.key":      "value2",
 				"tokenizer.key":           "value3",
 				"adapter.key":             "value4",
-			}, ff.KV()); diff != "" {
+			}, ff.KV(), cmp.AllowUnexported(array[int32]{}, array[int64]{})); diff != "" {
 				t.Errorf("Mismatch (-want +got):\n%s", diff)
 			}
 
 			if diff := cmp.Diff(Tensors{
-				Offset: 800,
+				Offset: 992,
 				items: []*Tensor{
 					{Name: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
 					{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},

integration/llm_image_test.go

@@ -33,6 +33,9 @@ func TestVisionModels(t *testing.T) {
 			// Qwen 3 VL mixture of experts
 			model: "qwen3-vl:30b",
 		},
+		{
+			model: "ministral-3",
+		},
 	}
 
 	for _, v := range testCases {

integration/tools_test.go

@@ -30,6 +30,7 @@ func TestAPIToolCalling(t *testing.T) {
 		"mistral":       6,
 		"qwen2.5":       6,
 		"qwen2":         6,
+		"ministral-3":   20,
 		"mistral-nemo":  9,
 		"mistral-small": 16,
 		"mixtral:8x22b": 80,

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
 		"gemma3n:e2b",
@@ -167,6 +168,7 @@ var (
 		"medllama2",
 		"megadolphin",
 		"minicpm-v",
+		"ministral-3",
 		"mistral-large",
 		"mistral-nemo",
 		"mistral-openorca",
@@ -270,6 +272,7 @@ var (
 		"mistral",
 		"qwen2.5",
 		"qwen2",
+		"ministral-3",
 		"mistral-nemo",
 		"mistral-small",
 		"mixtral:8x22b",

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
+char const *LLAMA_COMMIT = "7f8ef50cce40e3e7e4526a3696cb45658190e69a";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -22,6 +22,9 @@ include /src/llama.*
 include /src/llama-*.*
 include /src/unicode-data.*
 include /src/unicode.*
+include /src/models/
+include /src/models/*.h
+include /src/models/*.cpp
 include /vendor/
 include /vendor/miniaudio/
 include /vendor/miniaudio/*.h

llama/llama.cpp/common/common.cpp

@@ -8,6 +8,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -26,7 +27,6 @@
 #include <sstream>
 #include <string>
 #include <thread>
-#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,6 +60,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+common_time_meas::~common_time_meas() {
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
+    }
+}
+
 //
 // CPU utils
 //
@@ -355,11 +363,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
-        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
-            common_log_add(common_log_main(), level, "%s", text);
-        }
-    }, NULL);
+    llama_log_set(common_log_default_callback, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -908,11 +912,96 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
+std::vector<common_file_info> fs_list_files(const std::string & path) {
+    std::vector<common_file_info> files;
+    if (path.empty()) return files;
+
+    std::filesystem::path dir(path);
+    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
+        return files;
+    }
+
+    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
+        try {
+            // Only include regular files (skip directories)
+            const auto & p = entry.path();
+            if (std::filesystem::is_regular_file(p)) {
+                common_file_info info;
+                info.path = p.string();
+                info.name = p.filename().string();
+                try {
+                    info.size = static_cast<size_t>(std::filesystem::file_size(p));
+                } catch (const std::filesystem::filesystem_error &) {
+                    info.size = 0;
+                }
+                files.push_back(std::move(info));
+            }
+        } catch (const std::filesystem::filesystem_error &) {
+            // skip entries we cannot inspect
+            continue;
+        }
+    }
+
+    return files;
+}
+
 
 //
 // Model utils
 //
 
+static inline void common_init_sampler_from_model(
+    const llama_model * model,
+    common_params_sampling & sparams) {
+
+    const uint64_t config = sparams.user_sampling_config;
+
+    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[64] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            int32_t v = strtol(buf, &end, 10);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[128] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            float v = strtof(buf, &end);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    // Sampling sequence
+    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
+        char buf[512] = {0};
+        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
+            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
+            if (!sampler_names.empty()) {
+                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            }
+        }
+    }
+
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
+}
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
@@ -924,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    common_init_sampler_from_model(model, params.sampling);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     auto cparams = common_context_params_to_llama(params);

llama/llama.cpp/common/common.h

@@ -2,17 +2,15 @@
 
 #pragma once
 
+#include "ggml-opt.h"
+#include "llama-cpp.h"
+
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
-#include <sstream>
-#include <cmath>
-
-#include "ggml-opt.h"
-#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -30,6 +28,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct common_time_meas {
+    common_time_meas(int64_t & t_acc, bool disable = false);
+    ~common_time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
 struct common_adapter_lora_info {
     std::string path;
     float scale;
@@ -133,6 +140,22 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
+enum common_params_sampling_config : uint64_t {
+    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
+    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
+};
+
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -165,6 +188,8 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
+    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
+
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 
@@ -406,6 +431,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -458,7 +485,8 @@ struct common_params {
     float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
-    bool is_pp_shared = false;
+    bool is_pp_shared   = false;
+    bool is_tg_separate = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -505,6 +533,10 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+
+    bool has_speculative() const {
+        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
+    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -605,6 +637,13 @@ bool fs_create_directory_with_parents(const std::string & path);
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
+struct common_file_info {
+    std::string path;
+    std::string name;
+    size_t      size = 0; // in bytes
+};
+std::vector<common_file_info> fs_list_files(const std::string & path);
+
 //
 // Model utils
 //

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -303,6 +303,8 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
+std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
+
 class SchemaConverter {
 private:
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
@@ -601,7 +603,10 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        auto it = ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
+        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -774,11 +779,24 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
+                            if (target.is_object() && target.contains(sel)) {
+                                target = target[sel];
+                            } else if (target.is_array()) {
+                                size_t sel_index;
+                                try {
+                                    sel_index = std::stoul(sel);
+                                } catch (const std::invalid_argument & e) {
+                                    sel_index = target.size();
+                                }
+                                if (sel_index >= target.size()) {
+                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                    return;
+                                }
+                                target = target[sel_index];
+                            } else {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
-                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }

llama/llama.cpp/common/json-schema-to-grammar.h

@@ -18,4 +18,6 @@ struct common_grammar_options {
     bool dotall = false;
 };
 
+std::string gbnf_format_literal(const std::string & literal);
+
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

llama/llama.cpp/common/log.cpp

@@ -442,3 +442,9 @@ void common_log_set_prefix(struct common_log * log, bool prefix) {
 void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
+
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
+    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+        common_log_add(common_log_main(), level, "%s", text);
+    }
+}

llama/llama.cpp/common/log.h

@@ -36,6 +36,8 @@ extern int common_log_verbosity_thold;
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
+
 // the common_log uses an internal worker thread to print/write log messages
 // when the worker thread is paused, incoming log messages are discarded
 struct common_log;

llama/llama.cpp/common/sampling.cpp

@@ -3,9 +3,10 @@
 #include "common.h"
 #include "log.h"
 
-#include <cmath>
-#include <unordered_map>
 #include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <unordered_map>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -112,6 +113,13 @@ struct common_sampler {
 
     llama_token_data_array cur_p;
 
+    void reset() {
+        prev.clear();
+
+        llama_sampler_reset(grmr);
+        llama_sampler_reset(chain);
+    }
+
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
@@ -128,6 +136,12 @@ struct common_sampler {
 
         cur_p = { cur.data(), cur.size(), -1, false };
     }
+
+    common_time_meas tm() {
+        return common_time_meas(t_total_us, params.no_perf);
+    }
+
+    mutable int64_t t_total_us = 0;
 };
 
 std::string common_params_sampling::print() const {
@@ -298,6 +312,8 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 }
 
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    const auto tm = gsmpl->tm();
+
     if (accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -308,9 +324,7 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
-    llama_sampler_reset(gsmpl->grmr);
-
-    llama_sampler_reset(gsmpl->chain);
+    gsmpl->reset();
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
@@ -327,16 +341,54 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 
+    const double t_sampling_ms = gsmpl ? 1e-3*gsmpl->t_total_us : 0;
+
+    llama_perf_sampler_data data_smpl;
+    llama_perf_context_data data_ctx;
+
+    memset(&data_smpl, 0, sizeof(data_smpl));
+    memset(&data_ctx,  0, sizeof(data_ctx));
+
     if (gsmpl) {
-        llama_perf_sampler_print(gsmpl->chain);
+        auto & data = data_smpl;
+
+        data = llama_perf_sampler(gsmpl->chain);
+
+        // note: the sampling time includes the samplers time + extra time spent in common/sampling
+        LOG_INF("%s:    sampling time = %10.2f ms\n", __func__, t_sampling_ms);
+        LOG_INF("%s:    samplers time = %10.2f ms / %5d tokens\n", __func__, data.t_sample_ms, data.n_sample);
     }
+
     if (ctx) {
-        llama_perf_context_print(ctx);
+        auto & data = data_ctx;
+
+        data = llama_perf_context(ctx);
+
+        const double t_end_ms = 1e-3 * ggml_time_us();
+
+        const double t_total_ms = t_end_ms - data.t_start_ms;
+        const double t_unacc_ms = t_total_ms - (t_sampling_ms + data.t_p_eval_ms + data.t_eval_ms);
+        const double t_unacc_pc = 100.0 * t_unacc_ms /  t_total_ms;
+
+        LOG_INF("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
+        LOG_INF("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
+        LOG_INF("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
+        LOG_INF("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+        LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
+        LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
+
         llama_memory_breakdown_print(ctx);
     }
 }
 
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    llama_synchronize(ctx);
+
+    // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
+    const auto tm = gsmpl->tm();
+
     gsmpl->set_logits(ctx, idx);
 
     auto & grmr  = gsmpl->grmr;
@@ -428,6 +480,8 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 // helpers
 
 llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    const auto tm = gsmpl->tm();
+
     auto * res = &gsmpl->cur_p;
 
     if (do_sort && !res->sorted) {

llama/llama.cpp/include/llama.h

@@ -83,6 +83,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_NORM   = 0,
         LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
         LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_IMROPE = GGML_ROPE_TYPE_IMROPE,
         LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
@@ -245,6 +246,21 @@ extern "C" {
         LLAMA_KV_OVERRIDE_TYPE_STR,
     };
 
+    enum llama_model_meta_key {
+        LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_K,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIN_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD,
+        LLAMA_MODEL_META_KEY_SAMPLING_TEMP,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA,
+    };
+
     struct llama_model_kv_override {
         enum llama_model_kv_override_type tag;
 
@@ -460,7 +476,11 @@ extern "C" {
     LLAMA_API bool llama_supports_gpu_offload(void);
     LLAMA_API bool llama_supports_rpc        (void);
 
+    // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions
+    //       In some cases the requested values via llama_context_params may differ from the actual values used by the context
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ctx_seq  (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
@@ -481,6 +501,7 @@ extern "C" {
 
     LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
@@ -512,6 +533,9 @@ extern "C" {
     // Get the number of metadata key/value pairs
     LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
 
+    // Get sampling metadata key name. Returns nullptr if the key is invalid
+    LLAMA_API const char * llama_model_meta_key_str(enum llama_model_meta_key key);
+
     // Get metadata key name by index
     LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
 
@@ -584,7 +608,7 @@ extern "C" {
     LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
 
     // Manually free a LoRA adapter
-    // Note: loaded adapters will be free when the associated model is deleted
+    // NOTE: loaded adapters will be free when the associated model is deleted
     LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
 
     // Get the invocation tokens if the current lora is an alora
@@ -1110,8 +1134,6 @@ extern "C" {
     //        // sample from the logits of the last token in the batch
     //        const llama_token id = llama_sampler_sample(smpl, ctx, -1);
     //
-    //        // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.)
-    //        llama_sampler_accept(smpl, id);
     //        ...
     //    }
     //

llama/llama.cpp/src/llama-arch.cpp

@@ -32,6 +32,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
     { LLM_ARCH_QWEN3,            "qwen3"            },
     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
+    { LLM_ARCH_QWEN3NEXT,        "qwen3next"        },
+    { LLM_ARCH_QWEN3VL,          "qwen3vl"          },
+    { LLM_ARCH_QWEN3VLMOE,       "qwen3vlmoe"       },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -89,6 +92,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
+    { LLM_ARCH_AFMOE,            "afmoe"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
@@ -104,23 +108,39 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_APERTUS,          "apertus"          },
+    { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
+    { LLM_ARCH_COGVLM,           "cogvlm"           },
+    { LLM_ARCH_RND1,             "rnd1"             },
+    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-    { LLM_KV_GENERAL_TYPE,                 "general.type"                          },
-    { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
-    { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
-    { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
-    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
-    { LLM_KV_GENERAL_NAME,                 "general.name"                          },
-    { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
-    { LLM_KV_GENERAL_VERSION,              "general.version"                       },
-    { LLM_KV_GENERAL_URL,                  "general.url"                           },
-    { LLM_KV_GENERAL_DESCRIPTION,          "general.description"                   },
-    { LLM_KV_GENERAL_LICENSE,              "general.license"                       },
-    { LLM_KV_GENERAL_SOURCE_URL,           "general.source.url"                    },
-    { LLM_KV_GENERAL_SOURCE_HF_REPO,       "general.source.huggingface.repository" },
+    { LLM_KV_GENERAL_TYPE,                     "general.type"                          },
+    { LLM_KV_GENERAL_ARCHITECTURE,             "general.architecture"                  },
+    { LLM_KV_GENERAL_QUANTIZATION_VERSION,     "general.quantization_version"          },
+    { LLM_KV_GENERAL_ALIGNMENT,                "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,                "general.file_type"                     },
+    { LLM_KV_GENERAL_SAMPLING_SEQUENCE,        "general.sampling.sequence"             },
+    { LLM_KV_GENERAL_SAMPLING_TOP_K,           "general.sampling.top_k"                },
+    { LLM_KV_GENERAL_SAMPLING_TOP_P,           "general.sampling.top_p"                },
+    { LLM_KV_GENERAL_SAMPLING_MIN_P,           "general.sampling.min_p"                },
+    { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability"      },
+    { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,   "general.sampling.xtc_threshold"        },
+    { LLM_KV_GENERAL_SAMPLING_TEMP,            "general.sampling.temp"                 },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,  "general.sampling.penalty_last_n"       },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,  "general.sampling.penalty_repeat"       },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT,        "general.sampling.mirostat"             },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,    "general.sampling.mirostat_tau"         },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,    "general.sampling.mirostat_eta"         },
+    { LLM_KV_GENERAL_NAME,                     "general.name"                          },
+    { LLM_KV_GENERAL_AUTHOR,                   "general.author"                        },
+    { LLM_KV_GENERAL_VERSION,                  "general.version"                       },
+    { LLM_KV_GENERAL_URL,                      "general.url"                           },
+    { LLM_KV_GENERAL_DESCRIPTION,              "general.description"                   },
+    { LLM_KV_GENERAL_LICENSE,                  "general.license"                       },
+    { LLM_KV_GENERAL_SOURCE_URL,               "general.source.url"                    },
+    { LLM_KV_GENERAL_SOURCE_HF_REPO,           "general.source.huggingface.repository" },
 
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
@@ -146,6 +166,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
+    { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -329,6 +350,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_AFMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_GATE,       "blk.%d.attn_gate" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -781,6 +832,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_QWEN3NEXT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_SSM_A,              "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_CONV1D,         "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,             "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_BETA_ALPHA,     "blk.%d.ssm_ba" },
+            { LLM_TENSOR_SSM_IN,             "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_NORM,           "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,            "blk.%d.ssm_out" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VLMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -2168,7 +2290,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
@@ -2190,7 +2312,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_FFN_GATE_INP,      "blk.%d.ffn_gate_inp" },
             { LLM_TENSOR_FFN_GATE_EXPS,     "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,     "blk.%d.ffn_down_exps" },
@@ -2332,6 +2454,84 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
         },
     },
+    {
+        LLM_ARCH_MINIMAX_M2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
+    {
+        LLM_ARCH_PANGU_EMBED,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_COGVLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" },
+            { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" },
+            { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
+            { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
+            { LLM_TENSOR_VISEXP_FFN_UP,   "blk.%d.vis_up" },
+        },
+    },
+    {
+        LLM_ARCH_RND1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2340,11 +2540,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
     },
 };
 
+// declare information about the model weight tensors:
+// - the layer in which the tensor is going to be used. this is needed in order to assign the correct buffer type for the weight
+// - the operator which is going to use the weight. this is needed to determine if the respective backend supports the operator
+//
+// for example, input layers are usually assigned to CPU/host buffer types
+//
+// a mismatch between the declared information and the actual layer/op in which the tensor is used can lead to sub-optimal
+//   assignment of the buffer types and extra overhead during computation
+// example: https://github.com/ggml-org/llama.cpp/pull/17548
+//
 static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TOKEN_EMBD,                 {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_POS_EMBD,                   {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_TOKEN_TYPES,                {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2361,6 +2571,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_GATE,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2398,6 +2609,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_X,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_DT,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SSM_BETA_ALPHA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2509,6 +2721,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_QKV,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_OUT,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
     {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2592,6 +2809,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_QWEN3NEXT:
             return true;
         default:
             return false;
@@ -2603,6 +2821,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             return true;
         default:
             return false;

llama/llama.cpp/src/llama-arch.h

@@ -36,6 +36,9 @@ enum llm_arch {
     LLM_ARCH_QWEN2VL,
     LLM_ARCH_QWEN3,
     LLM_ARCH_QWEN3MOE,
+    LLM_ARCH_QWEN3NEXT,
+    LLM_ARCH_QWEN3VL,
+    LLM_ARCH_QWEN3VLMOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -93,6 +96,7 @@ enum llm_arch {
     LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
+    LLM_ARCH_AFMOE,
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,
@@ -108,6 +112,10 @@ enum llm_arch {
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_GROVEMOE,
     LLM_ARCH_APERTUS,
+    LLM_ARCH_MINIMAX_M2,
+    LLM_ARCH_COGVLM,
+    LLM_ARCH_RND1,
+    LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -117,6 +125,18 @@ enum llm_kv {
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
     LLM_KV_GENERAL_FILE_TYPE,
+    LLM_KV_GENERAL_SAMPLING_SEQUENCE,
+    LLM_KV_GENERAL_SAMPLING_TOP_K,
+    LLM_KV_GENERAL_SAMPLING_TOP_P,
+    LLM_KV_GENERAL_SAMPLING_MIN_P,
+    LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,
+    LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,
+    LLM_KV_GENERAL_SAMPLING_TEMP,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -150,6 +170,7 @@ enum llm_kv {
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_NEXTN_PREDICT_LAYERS,
+    LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -308,6 +329,7 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_POST_NORM,
     LLM_TENSOR_ATTN_ROT_EMBD,
     LLM_TENSOR_ATTN_SINKS,
+    LLM_TENSOR_ATTN_GATE,
     LLM_TENSOR_FFN_GATE_INP,
     LLM_TENSOR_FFN_GATE_INP_SHEXP,
     LLM_TENSOR_FFN_NORM,
@@ -362,6 +384,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_SSM_BETA_ALPHA,      // qwen3next
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
@@ -458,6 +481,11 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_VISEXP_ATTN_QKV,
+    LLM_TENSOR_VISEXP_ATTN_OUT,
+    LLM_TENSOR_VISEXP_FFN_GATE,
+    LLM_TENSOR_VISEXP_FFN_DOWN,
+    LLM_TENSOR_VISEXP_FFN_UP,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,

llama/llama.cpp/src/llama-batch.cpp

@@ -215,6 +215,7 @@ bool llama_batch_allocr::init(
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
             /*.n_seqs_unq   =*/ (uint32_t) this->seq_id_unq.size(),
+            /*.n_pos        =*/ n_pos_per_embd,
             /*.token        =*/ batch.token,
             /*.embd         =*/ batch.embd,
             /*.pos          =*/ batch.pos,
@@ -251,46 +252,72 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (uint32_t s = 0; s < n_seq_max; ++s) {
-        if (seq_pos[s].empty()) {
-            continue;
-        }
+    if (n_pos_per_embd > 1) {
+        // M-RoPE case: allow position to "jump" forward only (non-continuous positions are allowed)
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
 
-        const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
-
-        if (p0 >= 0) {
-            bool ok = true;
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
 
             if (batch.token) {
+                if (p0 >= 0 && p0 >= seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X < Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            } else {
+                // embedding inputs can have overlapping positions
+                if (p0 >= 0 && p0 > seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X <= Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            }
+        }
+    } else {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
+
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
+
+            if (p0 >= 0) {
+                bool ok = true;
+
                 if (seq_pos_min(s) != p0 + 1) {
                     ok = false;
                 }
-            } else {
-                assert(batch.embd);
 
-                // for embeddings (typically used as vision input), we allow them to have repeating positions
-                // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
-                if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) {
-                    ok = false;
+                if (!ok) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " it is required that the sequence positions remain consecutive: Y = X + 1\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
                 }
             }
 
-            if (!ok) {
-                LLAMA_LOG_ERROR(
-                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
-                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
-                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
-                        " it is required that the sequence positions remain consecutive: Y = X + 1\n",
-                        __func__, s, s, p0, s, seq_pos_min(s));
-
+            if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+                LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
                 return false;
             }
         }
-
-        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
-            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
-            return false;
-        }
     }
 
     if (memory) {
@@ -389,6 +416,7 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
@@ -655,10 +683,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
-
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
-    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
+    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_per_embd;
 
     udata->token     .resize(n_tokens);
     udata->embd      .resize(n_embd_all);
@@ -680,8 +706,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
             memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
-        for (int j = 0; j < n_pos_cur; ++j) {
-            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+        for (size_t j = 0; j < (size_t)n_pos_per_embd; ++j) {
+            // if we are using M-RoPE
+            //     if the current batch is text, we need to broadcast the same position across all RoPE sections
+            //     otherwise, the input batch is image embeddings, we copy the positions as-is
+            // if we are not using M-RoPE, there is only one position per token (this loop runs only once)
+            size_t src_off = batch.token ? 0 : j*batch.n_tokens;
+            udata->pos[j*n_tokens + i] = batch.pos[src_off + idxs[i]];
         }
 
         udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
@@ -710,6 +741,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ batch.token ? udata->token.data() : nullptr,
         /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,

llama/llama.cpp/src/llama-batch.h

@@ -17,6 +17,16 @@ struct llama_ubatch {
         return b_equal_seqs != 0;
     }
 
+    // typical for M-RoPE cases:
+    //   0 - sequantial position of the tokens/embeddings in the sequence
+    //   1 - y position in the image
+    //   2 - x position in the image
+    //   3 - other
+    bool is_pos_2d() const {
+        // TODO @ngxson : we may need to check for model arch when more models use >1 positions
+        return n_pos >= 3;
+    }
+
     uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
                            //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
@@ -25,6 +35,7 @@ struct llama_ubatch {
     uint32_t n_seq_tokens; // tokens per sequence set
     uint32_t n_seqs;       // sequence sets in the ubatch
     uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
+    uint32_t n_pos;        // number of position inputs for each token/embedding
 
     // seq_id_unq: unique sequence ids in the ubatch
     // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
@@ -33,7 +44,7 @@ struct llama_ubatch {
     //                          // size               | idx | val
     llama_token  *  token;      // [n_tokens]         | i   | id, token
     float        *  embd;       // [n_embd, n_tokens] | i   | embd
-    llama_pos    *  pos;        // [n_tokens]         | i   | pos
+    llama_pos    *  pos;        // [n_tokens*n_pos]   | i   | pos
     int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
     llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id

llama/llama.cpp/src/llama-chat.cpp

@@ -73,6 +73,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
+    { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -213,6 +214,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_SEED_OSS;
     } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
         return LLM_CHAT_TEMPLATE_GROK_2;
+    } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
+        return LLM_CHAT_TEMPLATE_PANGU_EMBED;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -813,6 +816,35 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    }else if (tmpl == LLM_CHAT_TEMPLATE_PANGU_EMBED) {
+        // [unused9]系统：xxx[unused10]
+        // [unused9]用户：xxx[unused10]
+        // [unused9]助手：xxx[unused10]
+        // ...
+        for (size_t i = 0; i < chat.size(); ++i) {
+            const auto & msg = chat[i];
+            const std::string & role = msg->role;
+            const std::string & content = msg->content;
+
+            if (i == 0 && role != "system") {
+                ss << "[unused9]系统：[unused10]";
+            }
+
+            if (role == "system") {
+                ss << "[unused9]系统：" << content << "[unused10]";
+            } else if (role == "user") {
+                ss << "[unused9]用户：" << content << "[unused10]";
+            } else if (role == "assistant") {
+                ss << "[unused9]助手：" << content << "[unused10]";
+            } else if (role == "tool") {
+                ss << "[unused9]工具：" << content << "[unused10]";
+            } else if (role == "function") {
+                ss << "[unused9]方法：" << content << "[unused10]";
+            }
+        }
+        if (add_ass) {
+            ss << "[unused9]助手：";
+        }
     } else {
         // template not supported
         return -1;

llama/llama.cpp/src/llama-chat.h

@@ -53,6 +53,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
+    LLM_CHAT_TEMPLATE_PANGU_EMBED,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

llama/llama.cpp/src/llama-context.cpp

@@ -1,5 +1,6 @@
 #include "llama-context.h"
 
+#include "llama-arch.h"
 #include "llama-impl.h"
 #include "llama-batch.h"
 #include "llama-io.h"
@@ -21,6 +22,8 @@ llama_context::llama_context(
               llama_context_params params) :
     model(model),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
+    //     may need to be backend-dependent
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -112,11 +115,28 @@ llama_context::llama_context(
         }
     }
 
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
+
+    if (cparams.kv_unified) {
+        cparams.n_ctx_seq = cparams.n_ctx;
+    } else {
+        cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+        cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
+
+        if (cparams.n_ctx_seq == 0) {
+            throw std::runtime_error("n_ctx_seq == 0");
+        }
+
+        if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) {
+            cparams.n_ctx =  cparams.n_ctx_seq * cparams.n_seq_max;
+            LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx);
+        }
+    }
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
     LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_ctx_seq     = %u\n",   __func__, cparams.n_ctx_seq);
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
@@ -125,14 +145,14 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
     if (!hparams.vocab_only) {
@@ -268,9 +288,7 @@ llama_context::llama_context(
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
         }
-    }
 
-    if (!hparams.vocab_only) {
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -282,7 +300,7 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         // avoid reserving graphs with zero outputs - assume one output per sequence
@@ -343,7 +361,14 @@ llama_context::llama_context(
         {
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
-                throw std::runtime_error("failed to allocate compute pp buffers");
+                if (pipeline_parallel) {
+                    LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
+                    sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload));
+                    gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+                }
+                if (!gf) {
+                    throw std::runtime_error("failed to allocate compute pp buffers");
+                }
             }
 
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
@@ -448,8 +473,8 @@ uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
 
-uint32_t llama_context::n_ctx_per_seq() const {
-    return cparams.n_ctx / cparams.n_seq_max;
+uint32_t llama_context::n_ctx_seq() const {
+    return cparams.n_ctx_seq;
 }
 
 uint32_t llama_context::n_batch() const {
@@ -518,7 +543,7 @@ bool llama_context::memory_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -803,7 +828,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -972,7 +997,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
 
     const bool output_all = false;
 
@@ -1223,7 +1248,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
-        if (!sorted_output) {
+        if (!sorted_output && n_outputs > 1) {
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
@@ -1361,6 +1386,9 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes() const {
+    if (model.arch == LLM_ARCH_QWEN3NEXT) {
+        return std::max<uint32_t>(8192u, 32u*model.n_tensors());
+    }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
@@ -2129,7 +2157,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2377,6 +2405,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) {
     return ctx->n_ctx();
 }
 
+uint32_t llama_n_ctx_seq(const llama_context * ctx) {
+    return ctx->n_ctx_seq();
+}
+
 uint32_t llama_n_batch(const llama_context * ctx) {
     return ctx->n_batch();
 }

llama/llama.cpp/src/llama-context.h

@@ -43,11 +43,11 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    uint32_t n_ctx()         const;
-    uint32_t n_ctx_per_seq() const;
-    uint32_t n_batch()       const;
-    uint32_t n_ubatch()      const;
-    uint32_t n_seq_max()     const;
+    uint32_t n_ctx()     const;
+    uint32_t n_ctx_seq() const;
+    uint32_t n_batch()   const;
+    uint32_t n_ubatch()  const;
+    uint32_t n_seq_max() const;
 
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;

llama/llama.cpp/src/llama-cparams.h

@@ -8,6 +8,7 @@
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
+    uint32_t n_ctx_seq;       // context for a single sequence
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;

llama/llama.cpp/src/llama-grammar.cpp

@@ -6,8 +6,10 @@
 
 #include <cmath>
 #include <algorithm>
+#include <cstdint>
 #include <stdexcept>
 
+#define MAX_REPETITION_THRESHOLD 2000
 //
 // helpers
 //
@@ -345,8 +347,10 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-    auto handle_repetitions = [&](int min_times, int max_times) {
-
+    // use UINT64_MAX as the empty value because we aligned to the proper uint64_t type so -1 can't be used
+    // (though it's technically the same as -1 now)
+    auto handle_repetitions = [&](uint64_t min_times, uint64_t max_times) {
+        bool no_max = max_times == UINT64_MAX;
         if (last_sym_start == rule.size()) {
             throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
         }
@@ -373,20 +377,20 @@ const char * llama_grammar_parser::parse_sequence(
             rule.resize(last_sym_start);
         } else {
             // Repeat the previous elements (min_times - 1) times
-            for (int i = 1; i < min_times; i++) {
+            for (uint64_t i = 1; i < min_times; i++) {
                 rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
 
         uint32_t last_rec_rule_id = 0;
-        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+        auto n_opt = no_max ? 1 : max_times - min_times;
 
         llama_grammar_rule rec_rule(prev_rule);
-        for (int i = 0; i < n_opt; i++) {
+        for (uint64_t i = 0; i < n_opt; i++) {
             rec_rule.resize(prev_rule.size());
             uint32_t rec_rule_id = generate_symbol_id( rule_name);
-            if (i > 0 || max_times < 0) {
-                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            if (i > 0 || no_max) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, no_max ? rec_rule_id : last_rec_rule_id});
             }
             rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
             rec_rule.push_back({LLAMA_GRETYPE_END, 0});
@@ -478,10 +482,10 @@ const char * llama_grammar_parser::parse_sequence(
                 throw std::runtime_error(std::string("expecting an int at ") + pos);
             }
             const char * int_end = parse_int(pos);
-            int min_times = std::stoul(std::string(pos, int_end - pos));
+            uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
             pos = parse_space(int_end, is_nested);
 
-            int max_times = -1;
+            uint64_t max_times = UINT64_MAX; // default: no max limit
 
             if (*pos == '}') {
                 max_times = min_times;
@@ -502,6 +506,10 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
+            bool has_max = max_times != UINT64_MAX;
+            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;

llama/llama.cpp/src/llama-graph.cpp

@@ -810,6 +810,9 @@ ggml_tensor * llm_graph_context::build_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
@@ -958,14 +961,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         // organize experts into n_expert_groups
         ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
 
-        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        ggml_tensor * group_scores = ggml_argsort_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
         group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
 
         // get top n_group_used expert groups
         group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
         group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
 
-        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        ggml_tensor * expert_groups = ggml_argsort_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
         cb(expert_groups, "ffn_moe_group_topk", il);
 
         // mask out the other groups
@@ -976,7 +979,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }
 
     // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
+    ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 
@@ -1006,10 +1009,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
         cb(weights_sum, "ffn_moe_weights_sum", il);
 
-        if (arch == LLM_ARCH_BAILINGMOE2) {
-            weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
-            cb(weights_sum, "ffn_moe_weights_sum_biased", il);
-        }
+        // Avoid division by zero, clamp to smallest number representable by F16
+        weights_sum = ggml_clamp(ctx0, weights_sum, 6.103515625e-5, INFINITY);
+        cb(weights_sum, "ffn_moe_weights_sum_clamped", il);
 
         weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
         cb(weights, "ffn_moe_weights_norm", il);
@@ -1091,6 +1093,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -1137,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd = hparams.n_embd_inp();
 
     auto inp = std::make_unique<llm_graph_input_embd>();
 
@@ -1274,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //    return cur;
     //}
 
-    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
+    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
     const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
@@ -1587,9 +1592,10 @@ ggml_tensor * llm_graph_context::build_attn(
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
 
     const auto * mctx_cur = inp->mctx;
 
@@ -2030,7 +2036,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck
 
     if (bidirectional) {
         relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
+        relative_position = std::abs(relative_position);
     } else {
         relative_position = -std::min<int32_t>(relative_position, 0);
     }

llama/llama.cpp/src/llama-hparams.cpp

@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_embd_inp() const {
+    uint32_t n_embd_inp = n_embd;
+
+    if (n_deepstack_layers > 0) {
+        n_embd_inp += n_embd * n_deepstack_layers;
+    }
+
+    return n_embd_inp;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
@@ -148,7 +158,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {
-    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
 }
 
 bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {

llama/llama.cpp/src/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
+#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
@@ -185,6 +185,9 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // qwen3vl deepstack
+    uint32_t n_deepstack_layers = 0;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -226,6 +229,9 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    // dimension of main + auxiliary input embeddings
+    uint32_t n_embd_inp() const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

llama/llama.cpp/src/llama-impl.cpp

@@ -20,10 +20,10 @@ static llama_logger_state g_logger_state;
 time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
 
 time_meas::~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += ggml_time_us() - t_start_us;
-        }
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
     }
+}
 
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);

llama/llama.cpp/src/llama-kv-cache-iswa.cpp

@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {

llama/llama.cpp/src/llama-kv-cache.cpp

@@ -8,6 +8,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cmath>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -37,8 +38,15 @@ llama_kv_cache::llama_kv_cache(
 
     const uint32_t n_layer_kv = hparams.n_layer_kv();
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -53,13 +61,12 @@ llama_kv_cache::llama_kv_cache(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     GGML_ASSERT(n_stream == 1 || n_stream == n_seq_max);
@@ -167,11 +174,8 @@ llama_kv_cache::llama_kv_cache(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -179,7 +183,7 @@ llama_kv_cache::llama_kv_cache(
         LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
 
         ggml_backend_buffer_clear(buf, 0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -203,7 +207,7 @@ void llama_kv_cache::clear(bool data) {
     }
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -334,6 +338,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             llama_pos pos   = v_cells[s0].pos_get(i);
             llama_pos shift = v_cells[s0].get_shift(i);
 
+            llama_kv_cell_ext ext = v_cells[s0].ext_get(i);
+
             if (shift != 0) {
                 pos -= shift;
                 assert(pos >= 0);
@@ -345,6 +351,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             if (shift != 0) {
                 v_cells[s1].pos_add(i, shift);
             }
+
+            v_cells[s1].ext_set(i, ext);
         }
     }
 
@@ -379,6 +387,7 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
 
 void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
     auto & head  = v_heads[seq_to_stream[seq_id]];
@@ -423,6 +432,7 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll
 
 void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
 
@@ -472,8 +482,8 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -896,6 +906,14 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
 
             cells.pos_set(idx, ubatch.pos[i]);
 
+            if (ubatch.is_pos_2d()) {
+                llama_kv_cell_ext ext {
+                    /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2],
+                    /*.y =*/ ubatch.pos[i + ubatch.n_tokens],
+                };
+                cells.ext_set(idx, ext);
+            }
+
             for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
                 cells.seq_add(idx, ubatch.seq_id[i][s]);
             }
@@ -957,10 +975,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;
@@ -1239,6 +1261,11 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
 
                 const llama_pos p1 = ubatch->pos[i];
 
+                // for M-RoPE
+                const bool is_2d = ubatch->is_pos_2d();
+                const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
+                const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
+
                 const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii);
 
                 for (uint32_t j = 0; j < n_kv; ++j) {
@@ -1258,6 +1285,14 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
                         continue;
                     }
 
+                    // M-RoPE causal mask
+                    if (causal_attn && is_2d && p0 == p1) {
+                        const auto & p0_ext = cells.ext_get(j);
+                        if (p0_ext.is_2d_gt(p1_x, p1_y)) {
+                            continue;
+                        }
+                    }
+
                     // apply SWA if any
                     if (is_masked_swa(p0, p1)) {
                         continue;
@@ -1298,7 +1333,7 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch
 size_t llama_kv_cache::total_size() const {
     size_t size = 0;
 
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
@@ -1340,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow  = cparams.yarn_beta_slow;
 
     const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
+    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
                                 // a normal RoPE should work, we just need to use the correct ordering
@@ -1551,6 +1586,9 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
             io.write(&pos,      sizeof(pos));
             io.write(&n_seq_id, sizeof(n_seq_id));
 
+            // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
+            //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
+
             for (const auto & seq_id : seq_ids) {
                 io.write(&seq_id, sizeof(seq_id));
             }
@@ -1696,6 +1734,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             return false;
         }
 
+        // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet
+        //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
         const auto head_cur = sinfo.head();
@@ -2010,8 +2050,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
-
-uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
-    // the FA kernels require padding to avoid extra runtime boundary checks
-    return cparams.flash_attn ? 256u : 32u;
-}

llama/llama.cpp/src/llama-kv-cache.h

@@ -19,8 +19,6 @@ struct llama_context;
 
 class llama_kv_cache : public llama_memory_i {
 public:
-    static uint32_t get_padding(const llama_cparams & cparams);
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -217,8 +215,8 @@ private:
     // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method

llama/llama.cpp/src/llama-kv-cells.h

@@ -5,9 +5,27 @@
 
 #include <bitset>
 #include <cassert>
-#include <vector>
-#include <set>
+#include <cstring>
 #include <map>
+#include <set>
+#include <vector>
+
+struct llama_kv_cell_ext {
+    // 2D spatial positions, typically used for M-RoPE
+    llama_pos x = 0;
+    llama_pos y = 0;
+
+    // return true if the current 2D spatial position is greater than other
+    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
+        return (y > oy) || (y == oy && x > ox);
+    }
+
+    void reset() {
+        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);
+
+        memset(this, 0, sizeof(*this));
+    }
+};
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -16,6 +34,7 @@ public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
             pos[i]   = -1;
+            ext[i].reset();
             shift[i] =  0;
             seq[i].reset();
         }
@@ -43,6 +62,7 @@ public:
 
     void resize(uint32_t n) {
         pos.resize(n);
+        ext.resize(n);
         shift.resize(n);
         seq.resize(n);
 
@@ -108,6 +128,7 @@ public:
             const auto idx = i + j;
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -126,6 +147,7 @@ public:
             const auto idx = idxs[j];
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -154,6 +176,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -184,6 +207,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -203,6 +227,7 @@ public:
         seq[i].reset();
 
         pos[i] = -1;
+        ext[i].reset();
         shift[i] = 0;
 
         used.erase(i);
@@ -221,6 +246,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -250,6 +276,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -340,6 +367,13 @@ public:
         return pos[i];
     }
 
+    const llama_kv_cell_ext & ext_get(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return ext[i];
+    }
+
     // note: call only if the cell is not empty
     llama_pos get_shift(uint32_t i) const {
         assert(i < pos.size());
@@ -368,6 +402,11 @@ public:
         used.insert(i);
     }
 
+    void ext_set(uint32_t i, llama_kv_cell_ext p) {
+        assert(i < ext.size());
+        ext[i] = p;
+    }
+
     // pos[i] = pos[i] + d
     // sets "has_shift" to true
     // note: call only if the cell is not empty
@@ -424,6 +463,9 @@ private:
 
     std::vector<llama_pos> pos;
 
+    // stores extra info per cell
+    std::vector<llama_kv_cell_ext> ext;
+
     // this array accumulates any applied shifts to the pos array since the last reset_shift() call
     // this is used to queue multiple updates to the pos array, which in the end can be applied in one go:
     //

llama/llama.cpp/src/llama-memory-recurrent.cpp

@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent(
     cells.clear();
     cells.resize(mem_size);
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     r_l.resize(n_layer);
@@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
         LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -129,7 +133,7 @@ void llama_memory_recurrent::clear(bool data) {
     used = 0;
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -147,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
     if (seq_id >= (int64_t) size) {
         // could be fatal
         return false;
@@ -156,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                 return false;
             }
@@ -364,8 +369,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -662,7 +667,7 @@ bool llama_memory_recurrent::get_can_shift() const {
 
 size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 

llama/llama.cpp/src/llama-memory-recurrent.h

@@ -109,8 +109,8 @@ private:
 
     const uint32_t n_seq_max = 1;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     size_t total_size() const;
 

llama/llama.cpp/src/llama-model.h

@@ -77,6 +77,7 @@ enum llm_type {
     LLM_TYPE_16B,
     LLM_TYPE_20B,
     LLM_TYPE_22B,
+    LLM_TYPE_26B,
     LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
@@ -113,8 +114,10 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5
@@ -234,6 +237,7 @@ struct llama_layer {
     struct ggml_tensor * wk_enc    = nullptr;
     struct ggml_tensor * wv_enc    = nullptr;
     struct ggml_tensor * wo_enc    = nullptr;
+    struct ggml_tensor * wqkv_gate = nullptr;
 
     // attention bias
     struct ggml_tensor * bq   = nullptr;
@@ -307,6 +311,9 @@ struct llama_layer {
     struct ggml_tensor * ssm_conv1d_b = nullptr;
     struct ggml_tensor * ssm_dt_b     = nullptr;
 
+    // qwen3next
+    struct ggml_tensor * ssm_beta_alpha = nullptr;
+
     // rwkv
     struct ggml_tensor * time_mix_w1         = nullptr;
     struct ggml_tensor * time_mix_w2         = nullptr;
@@ -385,6 +392,13 @@ struct llama_layer {
     // openai-moe
     struct ggml_tensor * attn_sinks = nullptr;
 
+    // cogvlm
+    struct ggml_tensor * visexp_attn_wqkv = nullptr;
+    struct ggml_tensor * visexp_attn_wo   = nullptr;
+    struct ggml_tensor * visexp_ffn_gate  = nullptr;
+    struct ggml_tensor * visexp_ffn_down  = nullptr;
+    struct ggml_tensor * visexp_ffn_up    = nullptr;
+
     // xIELU activation parameters for Apertus
     struct ggml_tensor * ffn_act_alpha_n = nullptr;
     struct ggml_tensor * ffn_act_alpha_p = nullptr;
@@ -503,9 +517,8 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;

llama/llama.cpp/src/llama-quant.cpp

@@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
                 // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
-                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
@@ -681,7 +681,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
-        } else if (remapped_name != it.first) {
+        }
+
+        if (remapped_name != it.first) {
             ggml_set_name(it.second.tensor, remapped_name.c_str());
             LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
         }
@@ -726,13 +728,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
-        int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
+        int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
         if (llama_model_has_encoder(&model)) {
-            // now n_attn_layer is the number of attention layers in the encoder
+            // now n_layer_attn is the number of attention layers in the encoder
             // for each decoder block, there are 2 attention layers
-            n_attn_layer += 2 * model.hparams.dec_n_layer;
+            n_layer_attn += 2 * model.hparams.dec_n_layer;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
+
+        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
+        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
+
+        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
+
+        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;

llama/llama.cpp/src/llama-sampling.cpp

@@ -4,6 +4,7 @@
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
+#include <array>
 #include <algorithm>
 #include <cassert>
 #include <cfloat>
@@ -471,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
     for (auto * smpl : chain->samplers) {
         llama_sampler_reset(smpl);
     }
-
-    chain->t_sample_us = 0;
-    chain->n_sample    = 0;
 }
 
 static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -1625,10 +1623,12 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        std::string trigger_pattern;
+        llama_grammar * grammar = nullptr;
         // TODO: remove trigger_words support.
         if (trigger_words != nullptr && num_trigger_words > 0) {
             GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
-            std::string trigger_pattern("[\\s\\S]*?(");
+            trigger_pattern = "[\\s\\S]*?(";
             for (size_t i = 0; i < num_trigger_words; ++i) {
                 static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
                 if (i > 0) {
@@ -1637,15 +1637,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            const auto * trigger_pattern_c = trigger_pattern.c_str();
-            trigger_patterns = &trigger_pattern_c;
-            num_trigger_patterns = 1;
+
+            std::array<const char *, 1> tmp_trigger_patterns = { trigger_pattern.c_str() };
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
+        } else {
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
         }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ grammar,
         };
         if (!ctx->grammar) {
             delete ctx;
@@ -2665,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
 void llama_perf_sampler_print(const struct llama_sampler * chain) {
     const auto data = llama_perf_sampler(chain);
 
-    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+    LLAMA_LOG_INFO("%s:    samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
 }
 
 void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2676,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
 
     auto * ctx = (struct llama_sampler_chain *) chain->ctx;
 
-    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_sample_us = 0;
+    ctx->n_sample    = 0;
 }

llama/llama.cpp/src/llama-vocab.cpp

@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -442,6 +443,17 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_AFMOE:
+                regex_exprs = {
+                    // Digit handling - uses custom implementation in unicode.cpp
+                    // Groups digits with leading 1-2 based on total length modulo 3
+                    "\\p{AFMoE_digits}",
+                    // CJK and Asian scripts (using direct Unicode literals)
+                    "[一-鿿㐀-䶿豈-﫿぀-ゟ゠-ヿ･-ﾟ⼀-⿟เ-๿຀-໿ក-៿က-႟ꩠ-ꩿꧠ-꧿가-힯ᄀ-ᇿ]+",
+                    // Main BPE pattern
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1012,7 +1024,7 @@ private:
         }
     private:
         uint32_t get_node(size_t index) {
-            if (index > xcda_array_size) {
+            if (index >= xcda_array_size) {
                 throw std::runtime_error("Index out of array bounds in XCDA array!");
             }
             return xcda_array[index];
@@ -1269,6 +1281,7 @@ struct llm_tokenizer_plamo2 : llm_tokenizer {
 
         // Build suffix list in lexicographical order of reversed strings
         std::vector<std::string> suffixes;
+        suffixes.reserve(suffix_to_score.size() + 1);
         for (const auto & pair : suffix_to_score) {
             suffixes.push_back(pair.first);
         }
@@ -1981,6 +1994,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "afmoe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_AFMOE;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;

llama/llama.cpp/src/llama-vocab.h

@@ -49,6 +49,8 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
+    LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
 };
 
 struct LLM_KV;

llama/llama.cpp/src/llama.go

@@ -5,4 +5,8 @@ package llama
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 // #cgo windows CPPFLAGS: -D_WIN32_WINNT=0x0602
 import "C"
-import _ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+
+import (
+	_ "github.com/ollama/ollama/llama/llama.cpp/src/models"
+	_ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+)

llama/llama.cpp/src/models/afmoe.cpp

@@ -0,0 +1,187 @@
+#include "models.h"
+
+llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // MuP scaling: embeddings * sqrt(hidden_size)
+    // mup_enabled = true, hidden_size = 1024, scale = 32.0
+    inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
+    cb(inpL, "inp_embd_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // dual attention normalization (pre)
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * attn_inp = cur;  // save input for gate computation
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            // compute gate from input
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Q/K normalization
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            // RoPE only for sliding_attention layers
+            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                                ((il + 1) % hparams.n_no_rope_layer_step) != 0;
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur_rope", il);
+            }
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,  // wo will be applied after gating
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
+            gate = ggml_sigmoid(ctx0, gate);
+            cb(gate, "attn_gate_sig", il);
+            cur = ggml_mul(ctx0, cur, gate);
+            cb(cur, "attn_gated", il);
+
+            // now apply output projection
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        // dual attention normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // dual ffn normalization (pre)
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE or dense FFN
+        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
+            // MoE layer with sigmoid routing, normalization, and scaling
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU,
+                    hparams.expert_weights_norm,           // norm_w (route_norm=True)
+                    hparams.expert_weights_scale,          // scale_w
+                    hparams.expert_weights_scale,          // w_scale (route_scale=2.826)
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // shared expert
+            if (hparams.n_expert_shared > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        } else {
+            // dense layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // dual ffn normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/apertus.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    auto *        inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/arcee.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // ARCEE uses relu^2 instead of silu
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/arctic.cpp

@@ -0,0 +1,138 @@
+#include "models.h"
+
+
+llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+        cb(ffn_out, "ffn_out", il);
+
+        // MoE
+        cur = build_norm(inpSA,
+                model.layers[il].ffn_norm_exps, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm_exps", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_out);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/arwkv7.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+
+llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/baichuan.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            switch (model.type) {
+                case LLM_TYPE_7B:
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    break;
+                case LLM_TYPE_13B:
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/bailingmoe.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    false, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/bailingmoe2.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+        cb(sa_out, "sa_out", il);
+
+        // MoE branch
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/bert.cpp

@@ -0,0 +1,176 @@
+#include "models.h"
+
+
+
+llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = nullptr;
+
+    if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+        inp_pos = build_inp_pos();
+    }
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+
+    // token types are hardcoded to zero ("Sentence A")
+    if (model.type_embd) {
+        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
+        inpL                    = ggml_add(ctx0, inpL, type_row0);
+    }
+    if (model.arch == LLM_ARCH_BERT) {
+        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
+    }
+    cb(inpL, "inp_embd", -1);
+
+    // embed layer norm
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // self-attention
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            } else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            // RoPE
+            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                model.arch == LLM_ARCH_JINA_BERT_V3) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // re-add the layer input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        // attention layer norm
+        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
+
+        if (model.layers[il].attn_norm_2 != nullptr) {
+            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input
+            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+            // MoE branch
+            cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
+                                model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
+                                LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+            cb(cur, "ffn_moe_out", il);
+        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                   model.arch == LLM_ARCH_JINA_BERT_V3) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // output layer norm
+        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/bitnet.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+
+llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].wq_scale) {
+                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
+            }
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            // B1.K
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].wk_scale) {
+                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
+            }
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            // B1.V
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].wv_scale) {
+                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
+            }
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+            cur = build_norm(cur,
+                    model.layers[il].attn_sub_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_sub_norm", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            if (model.layers[il].wo_scale) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
+            }
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward forward
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
+                NULL,                      NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_sub_out", il);
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_sub_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_sub_norm", il);
+
+        cur = build_lora_mm(model.layers[il].ffn_down, cur);
+        if (model.layers[il].ffn_down_scale) {
+            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
+        }
+        cb(cur, "ffn_down", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    // FIXME: do not use model.tok_embd directly, duplicate as model.output
+    cur = build_lora_mm(model.tok_embd, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/bloom.cpp

@@ -0,0 +1,101 @@
+#include "models.h"
+
+llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    inpL = build_norm(inpL,
+            model.tok_norm,
+            model.tok_norm_b,
+            LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/chameleon.cpp

@@ -0,0 +1,178 @@
+#include "models.h"
+
+#include <float.h>
+
+llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        if (hparams.swin_norm) {
+            cur = inpL;
+        } else {
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                        ggml_element_size(Qcur) * n_embd_head,
+                        ggml_element_size(Qcur) * n_embd_head * n_head,
+                        0);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm,
+                        model.layers[il].attn_q_norm_b,
+                        LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                        ggml_element_size(Kcur) * n_embd_head,
+                        ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                        0);
+                cb(Kcur, "Kcur", il);
+
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm,
+                        model.layers[il].attn_k_norm_b,
+                        LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (!hparams.swin_norm) {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output_with_img_logits", -1);
+
+    // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+    // Needs to be removed once image outputs are supported.
+    int img_token_end_idx = 8196;
+    int img_token_start_idx = 4;
+    int num_img_tokens = img_token_end_idx - img_token_start_idx;
+    // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+    // which ensures that text token values are always at least larger than image token values
+    ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+    img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+    cb(img_logits, "img_logits", -1);
+
+    cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/chatglm.cpp

@@ -0,0 +1,132 @@
+#include "models.h"
+
+
+llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+            }
+
+            //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+        }
+
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/codeshell.cpp

@@ -0,0 +1,111 @@
+#include "models.h"
+
+llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/cogvlm.cpp

@@ -0,0 +1,100 @@
+#include "models.h"
+
+llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor *inpL, *cur;
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    // check ubatch to see if we have input tokens (text)
+    // or an input embedding vector (image)
+    bool is_text;
+    if (ubatch.token) {
+        is_text = true;
+    } else {
+        is_text = false;
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        // get either the text or image weight tensors
+        ggml_tensor *wqkv, *wo;
+        ggml_tensor *ffn_gate, *ffn_down, *ffn_up;
+
+        if (is_text) {
+            wqkv     = model.layers[il].wqkv;
+            wo       = model.layers[il].wo;
+            ffn_gate = model.layers[il].ffn_gate;
+            ffn_down = model.layers[il].ffn_down;
+            ffn_up   = model.layers[il].ffn_up;
+        } else {
+            wqkv     = model.layers[il].visexp_attn_wqkv;
+            wo       = model.layers[il].visexp_attn_wo;
+            ffn_gate = model.layers[il].visexp_ffn_gate;
+            ffn_down = model.layers[il].visexp_ffn_down;
+            ffn_up   = model.layers[il].visexp_ffn_up;
+        }
+
+        ggml_tensor * inpSA = inpL;
+        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // build self attention
+        {
+            ggml_tensor * qkv = build_lora_mm(wqkv, cur);
+
+            // split qkv into Q, K, V along the first dimension
+            ggml_tensor * Qcur =
+                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], n_embd * ggml_element_size(qkv));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));
+
+            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);
+            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);
+
+            cur = build_attn(inp_attn,
+                wo, nullptr,
+                Qcur, Kcur, Vcur,
+                nullptr, nullptr, nullptr,
+                kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                ffn_up, NULL, NULL,
+                ffn_gate, NULL, NULL,
+                ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/cohere2-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_swa = hparams.is_swa(il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // rope freq factors for 128k context
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (is_swa) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/command-r.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+
+llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/dbrx.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+            cb(cur, "wqkv_clamped", il);
+
+            Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_out_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/deci.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA     = inpL;
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+        const int64_t n_head    = hparams.n_head(il);
+        const int64_t n_ff      = hparams.n_ff(il);
+
+        if (n_head == 0) {
+            // attention-free layer of Llama-3_1-Nemotron-51B
+            cur = inpL;
+        } else {
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        if (n_head > 0 && n_head_kv == 0) {
+            // "linear attention" of Llama-3_1-Nemotron-51B
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "wo", il);
+        } else if (n_head > 0) {
+            // self-attention
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
+        if (n_ff == 0) {
+            continue;
+        }
+        // modified to support attention-free layer of Llama-3_1-Nemotron-51B
+        ggml_tensor * ffn_inp = cur;
+        if (n_head > 0) {
+            ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+        }
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/deepseek.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+
+llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, hparams.expert_weights_scale,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/deepseek2.cpp

@@ -0,0 +1,237 @@
+#include "models.h"
+
+
+
+llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+    bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+
+    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    const float kq_scale    = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            if (!is_lite) {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                cb(q, "q", il);
+
+                q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+                cb(q, "q", il);
+
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                cb(q, "q", il);
+            } else {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(q, "q", il);
+            }
+            // split into {n_embd_head_qk_nope, n_head, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            if (is_mla) {
+                // {n_embd_head_qk_nope, n_tokens, n_head}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
+            } else {
+                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                cb(kv, "kv", il);
+
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * k_nope =
+                    ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);
+                cb(k_nope, "k_nope_view", il);
+
+                // and {n_embd_head_v, n_head, n_tokens}
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope));
+                cb(Vcur, "Vcur_view", il);
+
+                Vcur = ggml_cont(ctx0, Vcur);
+                cb(Vcur, "Vcur_cont", il);
+
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                cb(Kcur, "Kcur", il);
+
+                // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                cur = build_attn(inp_attn,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            }
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/dots1.cpp

@@ -0,0 +1,134 @@
+#include "models.h"
+
+
+
+llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/dream.cpp

@@ -0,0 +1,105 @@
+#include "models.h"
+
+
+
+llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    //copied from qwen2
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/ernie4-5-moe.cpp

@@ -0,0 +1,150 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        bool is_moe_layer =
+            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
+
+        if (!is_moe_layer) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                                        model.layers[il].ffn_gate_inp,
+                                        model.layers[il].ffn_up_exps,
+                                        model.layers[il].ffn_gate_exps,
+                                        model.layers[il].ffn_down_exps,
+                                        model.layers[il].ffn_exp_probs_b,
+                                        n_expert, n_expert_used,
+                                        LLM_FFN_SILU, true,
+                                        false, 0.0,
+                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                                        il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // Shared expert (if present)
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            } else {
+                cur = moe_out;
+            }
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/ernie4-5.cpp

@@ -0,0 +1,110 @@
+#include "models.h"
+
+llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/exaone.cpp

@@ -0,0 +1,114 @@
+#include "models.h"
+
+
+
+llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/exaone4.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+template <bool iswa>
+llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // use RoPE for SWA layers or non-SWA models
+        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
+
+        cur = inpL;
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_ffn(ffn_inp,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL, NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_exaone4<false>;
+template struct llm_build_exaone4<true>;

llama/llama.cpp/src/models/falcon-h1.cpp

@@ -0,0 +1,113 @@
+#include "models.h"
+
+
+
+llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Build the inputs in the recurrent & kv cache
+    auto * inp = build_inp_mem_hybrid();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cb(Qcur, "Qcur-post-rope", il);
+        cb(Kcur, "Kcur-post-rope", il);
+        cb(Vcur, "Vcur-post-rope", il);
+
+        ggml_tensor * attn_out = build_attn(inp->get_attn(),
+                                    model.layers[il].wo, NULL,
+                                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(attn_out, "attn_out", il);
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        // Mamba2 layer
+        cb(cur, "ssm_in", il);
+
+        ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        cb(ssm_out, "ssm_out", il);
+
+        // // Aggregation
+        cur   = ggml_add(ctx0, attn_out, ssm_out);
+        inpSA = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "layer_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = inpSA;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, inpSA);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/falcon.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * attn_norm;
+
+        attn_norm = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(attn_norm, "attn_norm", il);
+
+        // self-attention
+        {
+            if (model.layers[il].attn_norm_2) {
+                // Falcon-40B
+                cur = build_norm(inpL,
+                        model.layers[il].attn_norm_2,
+                        model.layers[il].attn_norm_2_b,
+                        LLM_NORM, il);
+                cb(cur, "attn_norm_2", il);
+            } else {
+                cur = attn_norm;
+            }
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            // using mode = 2 for neox mode
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
+            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
+            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+
+        // feed forward
+        {
+            cur = build_ffn(attn_norm, // !! use the attn norm, not the result
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/gemma-embedding.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+
+llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur =
+                build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/gemma.cpp

@@ -0,0 +1,112 @@
+#include "models.h"
+
+
+llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+            cb(Qcur, "Qcur_scaled", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/gemma2-iswa.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // final logit soft-capping
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+    cur = ggml_tanh(ctx0, cur);
+    cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/gemma3-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/gemma3n-iswa.cpp

@@ -0,0 +1,377 @@
+#include "models.h"
+
+
+
+llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model),
+    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_altup(model.hparams.n_embd_altup),
+    n_altup(model.hparams.n_altup),
+    i_altup_act(model.hparams.i_altup_act) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
+    ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
+
+    // inpL now has only 1 altup, project it to the rest of the altups
+    // these "added" altups will be concat to the last dim of inpL
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(inpL);
+        ggml_tensor * inp_repeated     = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
+        ggml_tensor * altup_added =
+            ggml_mul_mat(ctx0, model.altup_proj, inp_repeated);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_added);
+        altup_added                 = ggml_div(ctx0, ggml_mul(ctx0, altup_added, target_magnitude), new_magnitude);
+        inpL                        = ggml_concat(ctx0, inpL, altup_added, 2);  // shape: [n_embd, n_tokens, n_altup]
+        cb(inpL, "inp_stacked", -1);
+    }
+    // inpL now has shape:          [n_embd,       n_tokens, n_altup]
+    // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
+
+    for (int il = 0; il < n_layer; ++il) {
+        // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        ggml_tensor * cur         = inpL;                    // [n_embd, n_tokens, n_altup]
+        ggml_tensor * predictions = altup_predict(cur, il);  // [n_embd, n_tokens, n_altup]
+
+        // predicted value will go through self-attention and laurel
+        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);  // [n_embd, n_tokens]
+        cur                             = active_prediction;
+        cb(cur, "active_prediction", il);
+
+        // norm
+        cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // laurel
+        ggml_tensor * laurel_out = laurel(cur, il);  // [n_embd, n_tokens]
+
+        // self-attention
+        if (hparams.has_kv(il)) {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
+
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+            cb(Vcur, "Vcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+
+            cur = build_attn(inp_attn, model.layers[il].wo,
+                    NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+                    hparams.f_attention_scale, il);
+        } else {
+            // reuse KV cache of earlier layers
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(Qcur, "Qcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, active_prediction);  // [n_embd, n_tokens]
+        cb(cur, "attn_gated", il);
+
+        ggml_tensor * attn_laurel = ggml_scale(ctx0, ggml_add(ctx0, cur, laurel_out),
+                                               1.0f / sqrtf(2.0f));  // [n_embd, n_tokens]
+        cb(attn_laurel, "attn_laurel", il);
+
+        cur = build_norm(attn_laurel, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up, cur);
+            ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
+
+            if (il < n_layer_sparsity) {
+                // apply activation sparsity
+                gate_proj = gaussian_topk(gate_proj);
+            }
+            gate_proj = ggml_gelu(ctx0, gate_proj);
+
+            cur = ggml_mul(ctx0, up_proj, gate_proj);
+            cur = build_lora_mm(model.layers[il].ffn_down, cur);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", il);
+
+        ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel);  // [n_embd, n_tokens]
+        cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
+
+        ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il);  // [n_embd, n_tokens, n_altup]
+
+        ggml_tensor * first_prediction;                                                   // [n_embd, n_tokens]
+        {
+            first_prediction = view_2d_slice(corrected, i_altup_act);                     // [n_embd, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
+            first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
+            first_prediction = ggml_gelu(ctx0, first_prediction);                 // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_gated", il);
+            ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il);      // [n_embd_altup, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer);  // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_scaled", il);
+
+            first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction);  // [n_embd, n_tokens]
+            first_prediction =
+                build_norm(first_prediction, model.layers[il].per_layer_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(first_prediction, "first_prediction_out", il);
+        }
+        // equivalent to python code: corrected_predictions[1:] += first_prediction
+        {
+            ggml_tensor * slice_first = view_2d_slice(corrected, 0);
+            ggml_tensor * slice_rest  = ggml_view_3d(
+                ctx0, corrected, n_embd, n_tokens, n_altup - 1, ggml_row_size(corrected->type, n_embd),
+                ggml_row_size(corrected->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(corrected));
+            ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction);  // [n_embd, n_tokens, n_altup - 1]
+            corrected         = ggml_concat(ctx0, slice_first, tmp, 2);        // [n_embd, n_tokens, n_altup]
+        }
+        cur = corrected;                                                       // [n_embd, n_tokens, n_altup]
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;  // [n_embd, n_tokens, n_altup]
+
+    // cur now has multiple altup(s), we want to merge them back to 1 altup
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act));  // [n_embd, n_tokens]
+        // do a view to skip the first slice (active altup)
+        ggml_tensor * alt_slice =
+            ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, ggml_row_size(cur->type, n_embd),
+                         ggml_row_size(cur->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(cur));
+        ggml_tensor * altup_unembd =
+            ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
+        altup_unembd                = ggml_div(ctx0, ggml_mul(ctx0, altup_unembd, target_magnitude), new_magnitude);
+        cb(altup_unembd, "altup_unembd", -1);
+
+        // equivalent to torch.mean(hidden_states, dim=0)
+        cur = view_2d_slice(cur, 0);  // [n_embd, n_tokens]
+        for (int i = 0; i < n_altup - 1; ++i) {
+            cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
+        }
+        cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup));  // [n_embd, n_tokens]
+        cb(cur, "unembd_merged", -1);
+    }
+    // cur now has shape: [n_embd, n_tokens]
+
+    // TODO: move this to right after the last KV layer
+    {
+        // skip computing output for unused tokens
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+        cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    {
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) {
+    return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
+}
+
+// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
+ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
+    GGML_ASSERT(idx < (int) x->ne[2]);
+    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
+                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
+}
+
+// equivalent to get_per_layer_inputs() in python code
+// output shape: [n_embd_altup, n_layer, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
+    auto          inp = std::make_unique<llm_graph_input_embd>();
+    ggml_tensor * inp_per_layer;
+    if (ubatch.token) {
+        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+        ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
+        inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
+        inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
+        cb(inp_per_layer, "inp_per_layer_selected", -1);
+    } else {
+        GGML_ABORT("TODO: support embd input");
+    }
+    res->add_input(std::move(inp));
+    return inp_per_layer;
+}
+
+// equivalent to project_per_layer_inputs() in python code
+// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
+// output shape: [n_embd_altup, n_tokens, n_layer]
+ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
+    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
+    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
+
+    ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+    per_layer_proj               = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
+    per_layer_proj               = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
+    per_layer_proj               = build_norm(per_layer_proj, model.per_layer_proj_norm, NULL, LLM_NORM_RMS,
+                                              -1);  // [n_embd_altup, n_layer, n_tokens]
+    cb(per_layer_proj, "per_layer_proj", -1);
+
+    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
+    cb(inp_per_layer, "inp_per_layer", -1);
+
+    // permute to shape: [n_embd_altup, n_tokens, n_layer]
+    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
+    return inp_per_layer;
+}
+
+// input cur shape: [n_altup, n_tokens]
+// output    shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
+    ggml_tensor * tmp = cur;
+    tmp               = build_lora_mm(model.layers[il].laurel_l, tmp);
+    tmp               = build_lora_mm(model.layers[il].laurel_r, tmp);
+    tmp               = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
+    tmp               = ggml_add(ctx0, tmp, cur);
+    cb(tmp, "laurel_out", il);
+    return tmp;
+}
+
+// input x shape: [n_embd, n_tokens]
+// output  shape: [n_embd, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
+    ggml_tensor * mean = ggml_mean(ctx0, x);
+    ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
+                                                    1.0f / (float) (x->ne[0] - 1)));
+    ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
+    return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
+}
+
+//
+// altup functions
+//
+
+// equivalent to compute_router_modalities() in python code
+// input x shape: [n_embd,  n_tokens]
+// output  shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) {
+    ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il);
+
+    // router_input_scale
+    router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float) n_embd);
+
+    ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
+    return ggml_tanh(ctx0, output);  // [n_altup, n_tokens]
+}
+
+// input cur shape: [n_embd, n_tokens, n_altup]
+// output    shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
+    ggml_tensor * activated  = view_2d_slice(cur, i_altup_act);                 // [n_embd, n_tokens]
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
+    cb(all_coefs, "all_coefs", il);
+    // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
+    all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
+
+    // permute to [n_altup, n_embd, n_tokens]
+    ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+    ggml_tensor * predictions  = ggml_mul_mat(ctx0, cur_permuted, all_coefs);  // [n_altup, n_embd, n_tokens]
+
+    // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
+    predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
+    predictions = ggml_add(ctx0, predictions, cur);
+    cb(predictions, "predictions", il);
+
+    return predictions;
+}
+
+// input predictions       shape: [n_embd, n_tokens, n_altup]
+// input activated         shape: [n_embd, n_tokens]
+// output                  shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
+    ggml_tensor * innovation        = ggml_sub(ctx0, activated, active_prediction);  // [n_embd, n_tokens]
+    cb(innovation, "innovation", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities);  // [n_altup, n_tokens]
+    all_coefs               = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f);                    // + 1.0
+    cb(all_coefs, "all_coefs", il);
+    all_coefs = ggml_transpose(ctx0, all_coefs);                                               // [n_tokens, n_altup]
+    all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup);                           // [1, n_tokens, n_altup]
+
+    innovation              = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
+    ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs);   // [n_embd, n_tokens, n_altup]
+    corrected               = ggml_add(ctx0, corrected, predictions);  // [n_embd, n_tokens, n_altup]
+    cb(corrected, "corrected", il);
+
+    return corrected;
+}

llama/llama.cpp/src/models/glm4-moe.cpp

@@ -0,0 +1,153 @@
+#include "models.h"
+
+llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Only process up to last layer (skip final NextN layer)
+    // Final layer tensors are loaded but not processed in forward pass
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            }
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            }
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            }
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Apply Q/K norm if available (GLM-4.5 355B variant)
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+            }
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // Post-attention norm
+        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            // Dense FFN layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // Process routed experts using existing MoE infrastructure
+            ggml_tensor * routed_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(routed_out, "ffn_moe_out", il);
+
+            // Process shared expert on original input
+            ggml_tensor * shared_out = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(shared_out, "ffn_shexp_out", il);
+
+            // Final output: routed_output + shared_output
+            cur = ggml_add(ctx0, routed_out, shared_out);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/glm4.cpp

@@ -0,0 +1,127 @@
+#include "models.h"
+
+
+
+llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // Post-attention norm (new!)
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Add the input (residual connection after post-attention norm)
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            // Pre-MLP norm
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            // Post-MLP norm
+            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_mlp_norm", il);
+        }
+        // Add residual connection after post-MLP norm
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    // Final norm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}